[go: up one dir, main page]
Next Issue
Volume 24, September-1
Previous Issue
Volume 24, August-1
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
7.3
3.4
Journals
Sensors
Volume 24
Issue 16
sensors-logo

Journal Menu

► Journal Menu

Journal Browser

► Journal Browser
share announcement

Sensors, Volume 24, Issue 16 (August-2 2024) – 371 articles

Cover Story (view full-size image): Low Earth Orbit (LEO) satellites show significant potential for use in the global Internet of Things; however, the Doppler effect complicates Long Range (LoRa) uplink connectivity. This work assesses LEO satellite efficiency against Doppler effects using Artificial Intelligence techniques, tracking the Globalstar system and the International Space Station from Australia for 24 h starting 1 January 2024. We model the constellations, calculate latency and frequency offset, and develop a hybrid model to predict the Doppler profiles. The packet delivery exceeds 91% with optimal metadata settings. The model achieves 97.51% accuracy at Signal to Noise Ratios of 0–30 dB. Further research on LoRa Doppler effects and atmospheric factors is recommended. View this paper
  • Issues are regarded as officially published after their release is announced to the table of contents alert mailing list.
  • You may sign up for e-mail alerts to receive table of contents of newly released issues.
  • PDF is the official format for papers published in both, html and pdf forms. To view the papers in pdf format, click on the "PDF Full-text" link, and use the free Adobe Reader to open them.
Order results
Result details
Section
Show export options expand_more Show export options expand_less
Select all
Export citation of selected articles as:
19 pages, 3017 KiB  
Article
Adaptive Sensing Data Augmentation for Drones Using Attention-Based GAN
by Namkyung Yoon, Kiseok Kim, Sangmin Lee, Jin Hyoung Bai and Hwangnam Kim
Sensors 2024, 24(16), 5451; https://doi.org/10.3390/s24165451 - 22 Aug 2024
Cited by 1 | Viewed by 722
Abstract
Drones have become essential tools across various industries due to their ability to provide real-time data and perform automated tasks. However, integrating multiple sensors on a single drone poses challenges such as payload limitations and data management issues. This paper proposes a comprehensive [...] Read more.
Drones have become essential tools across various industries due to their ability to provide real-time data and perform automated tasks. However, integrating multiple sensors on a single drone poses challenges such as payload limitations and data management issues. This paper proposes a comprehensive system that leverages advanced deep learning techniques, specifically an attention-based generative adversarial network (GAN), to address data scarcity in drone-collected time-series sensor data. By adjusting sensing frequency based on operational conditions while maintaining data resolution, our system ensures consistent and high-quality data collection. The spatiotemporal The attention mechanism within the GAN enhances the generation of synthetic data, filling gaps caused by reduced sensing frequency with realistic data. This approach improves the efficiency and performance of various applications, such as precision agriculture, environmental monitoring, and surveillance. The experimental results demonstrated the effectiveness of our methodology in extending the operational range and duration of drones and providing reliable augmented data utilizing a variety of evaluation metrics. Furthermore, the superior performance of the proposed system was verified by comparing it with various comparative GAN models. Full article
(This article belongs to the Special Issue Internet of Things, Big Data and Smart Systems II)
►▼ Show Figures

Figure 1

Figure 1
<p>System Overview: Attention-based GAN for time-series sensor data augmentation.</p> Full article ">Figure 2
<p>Time Series Plots of Drone Sensor Data for ESP2 in <a href="#sec3dot2-sensors-24-05451" class="html-sec">Section 3.2</a>.</p> Full article ">Figure 3
<p>Exploratory data analysis on drone data for ESP2.</p> Full article ">Figure 4
<p>GAN architecture with a spatiotemporal attenion mechanism.</p> Full article ">Figure 5
<p>Distribution of each data variable over a total of 6 regions.</p> Full article ">Figure 6
<p>Discriminator training loss per datasets.</p> Full article ">Figure 7
<p>Accuracy by epoch of discriminator in the learning process.</p> Full article ">Figure 8
<p>Visualization of Real Data and Synthetic Data. Orange is synthetic data and blue is original data.</p> Full article ">Figure 9
<p>Comparison of drone voltage sensor data distribution in different regions. Orange is synthetic data and blue is original data.</p> Full article ">Figure 10
<p>Comparison of original voltage data and synthetic voltage data.</p> Full article ">Figure 11
<p>Comparison of baseline and augmented model with ESP2 dataset.</p> Full article ">
24 pages, 2169 KiB  
Article
Human Multi-Activities Classification Using mmWave Radar: Feature Fusion in Time-Domain and PCANet
by Yier Lin, Haobo Li and Daniele Faccio
Sensors 2024, 24(16), 5450; https://doi.org/10.3390/s24165450 - 22 Aug 2024
Viewed by 817
Abstract
This study introduces an innovative approach by incorporating statistical offset features, range profiles, time–frequency analyses, and azimuth–range–time characteristics to effectively identify various human daily activities. Our technique utilizes nine feature vectors consisting of six statistical offset features and three principal component analysis network [...] Read more.
This study introduces an innovative approach by incorporating statistical offset features, range profiles, time–frequency analyses, and azimuth–range–time characteristics to effectively identify various human daily activities. Our technique utilizes nine feature vectors consisting of six statistical offset features and three principal component analysis network (PCANet) fusion attributes. These statistical offset features are derived from combined elevation and azimuth data, considering their spatial angle relationships. The fusion attributes are generated through concurrent 1D networks using CNN-BiLSTM. The process begins with the temporal fusion of 3D range–azimuth–time data, followed by PCANet integration. Subsequently, a conventional classification model is employed to categorize a range of actions. Our methodology was tested with 21,000 samples across fourteen categories of human daily activities, demonstrating the effectiveness of our proposed solution. The experimental outcomes highlight the superior robustness of our method, particularly when using the Margenau–Hill Spectrogram for time–frequency analysis. When employing a random forest classifier, our approach outperformed other classifiers in terms of classification efficacy, achieving an average sensitivity, precision, F1, specificity, and accuracy of 98.25%, 98.25%, 98.25%, 99.87%, and 99.75%, respectively. Full article
(This article belongs to the Special Issue Radar and Multimodal Sensing for Ambient Assisted Living)
►▼ Show Figures

Figure 1

Figure 1
<p>This figure shows the range profiles of measured data with two subjects doing standing (lower) and falling (upper).</p> Full article ">Figure 2
<p>This figure displays the STFT image of measured data in <a href="#sensors-24-05450-f001" class="html-fig">Figure 1</a> with 800 Hz sampling frequency.</p> Full article ">Figure 3
<p>These images display the relationships between (<b>a</b>) horizontal and (<b>b</b>) vertical range and angles using elevation and azimuth information of measured data in <a href="#sensors-24-05450-f001" class="html-fig">Figure 1</a>.</p> Full article ">Figure 4
<p>This figure shows the instantaneous 3D point cloud of measured data in <a href="#sensors-24-05450-f003" class="html-fig">Figure 3</a>.</p> Full article ">Figure 5
<p>This figure shows the CNN-BiLSTM structure.</p> Full article ">Figure 6
<p>This figure displays the trainNetwork.</p> Full article ">Figure 7
<p>This figure shows the framework of our method.</p> Full article ">Figure 8
<p>This figure displays the experimental scene diagram and photo.</p> Full article ">Figure 9
<p>This figure shows the scatter plot of the fourteen categories of samples.</p> Full article ">Figure 10
<p>This figure shows the classification sensitivity of these combinations of actions using competitive temporal neural networks.</p> Full article ">Figure 11
<p>This figure shows our average sensitivity of fourteen categories via various classifiers under the time–frequency condition of STFT, Margenau–Hill Spectrogram, Choi–Williams, smoothed pseudo Wigner–Ville, and the joint of the above four methods.</p> Full article ">Figure 12
<p>This figure shows the accuracy of our method under different epoch numbers.</p> Full article ">Figure 13
<p>This figure shows the precision of our method (Test 6) and alternative method (Test 1–5).</p> Full article ">
17 pages, 34782 KiB  
Article
Non-Tectonic Geohazards of Guangdong Province, China, Monitored Using Sentinel-1A/B from 2015 to 2022
by Jincang Liu, Zhenhua Fu, Lipeng Zhou, Guangcai Feng, Yilin Wang and Wulinhong Luo
Sensors 2024, 24(16), 5449; https://doi.org/10.3390/s24165449 - 22 Aug 2024
Viewed by 517
Abstract
Guangdong Province, home to 21 cities and a permanent population of 127.06 million people, boasts the largest provincial economy in China, contributing 11.76% to the national GDP in 2023. However, it is prone to geological hazards due to its geological conditions, extreme weather, [...] Read more.
Guangdong Province, home to 21 cities and a permanent population of 127.06 million people, boasts the largest provincial economy in China, contributing 11.76% to the national GDP in 2023. However, it is prone to geological hazards due to its geological conditions, extreme weather, and extensive human activities. Geohazards not only endanger lives but also hinder regional economic development. Monitoring surface deformation regularly can promptly detect geological hazards and allow for effective mitigation strategies. Traditional ground subsidence monitoring methods are insufficient for comprehensive surveys and rapid monitoring of geological hazards in the whole province. Interferometric Synthetic Aperture Radar (InSAR) technology using satellite images can achieve wide-area geohazard monitoring. However, current geological hazard monitoring in Guangdong Province based on InSAR technology lacks regional analysis and statistics of surface deformation across the entire province. Furthermore, such monitoring fails to analyze the spatial–temporal characteristics of surface deformation and disaster evolution mechanisms by considering the local geological features. To address these issues, current work utilizes Sentinel-1A/B satellite data covering Guangdong Province from 2015 to 2022 to obtain the wide-area surface deformation in the whole province using the multi-temporal (MT) InSAR technology. Based on the deformation results, a wide-area deformation region automatic identification method is used to identify the surface deformation regions and count the deformation area in each city of Guangdong Province. By analyzing the results, we obtained the following findings: (1) Using the automatic identification algorithm we identified 2394 deformation regions. (2) Surface subsidence is concentrated in the delta regions and reclamation areas; over a 4 cm/year subsidence rate is observed in the hilly regions of northern Guangdong, particularly in mining areas. (3) Surface deformation is closely related to geological structures and human activities. (4) Sentinel-1 satellite C-band imagery is highly effective for wide-area geological hazard monitoring, but has limitations in monitoring small-area geological hazards. In the future, combining the high-spatial–temporal-resolution L-band imagery from the NISAR satellite with Sentinel-1 imagery will allow for comprehensive monitoring and early warning of geological hazards, achieving multiple geometric and platform perspectives for geological hazard monitoring and management in Guangdong Province. The findings of this study have significant reference value for the monitoring and management of geological disasters in Guangdong Province. Full article
(This article belongs to the Section Environmental Sensing)
►▼ Show Figures

Figure 1

Figure 1
<p>The spatial distribution, level, and types of geological hazards in Guangdong Province. (<b>a</b>) Geological background and disaster distribution of the study area on a color-shaded elevation map; (<b>b</b>) the pie chart of geological hazard level; (<b>c</b>) the pie chart of geological hazard types.</p> Full article ">Figure 2
<p>Coverage of the Sentinel-1 images over Guangdong Province from 2015 to 2022 including 11 frames in 5 tracks. The green line presents the footprint of Sentinel 1A/B data and the red line presents the boundary of Guangdong Province.</p> Full article ">Figure 3
<p>The interferogram network figures are based on spatial–temporal baselines.</p> Full article ">Figure 4
<p>Surface deformation velocity map of Guangdong Province for the period 2015–2022. A–F are six mining sites with large deformation; (<b>a</b>–<b>c</b>) are the zoom-ins of the Leizhou Peninsula, Pearl River Delta, and Hanjiang Delta.</p> Full article ">Figure 5
<p>The uncertainty of the surface deformation velocity map of Guangdong Province for the period 2015–2022.</p> Full article ">Figure 6
<p>Surface deformation velocity of the six selected mining areas in northern Guangdong Province. The figures (<b>a</b>–<b>f</b>) correspond to the six selected mining areas A–F in <a href="#sensors-24-05449-f004" class="html-fig">Figure 4</a>.</p> Full article ">Figure 7
<p>Surface deformation sequence results of (<b>a</b>) the Zhuhai reclamation area and (<b>b</b>) the uplift region of Puning, Jieyang.</p> Full article ">Figure 8
<p>Surface deformation rate map of coastal reclamation areas in Guangdong Province.</p> Full article ">Figure 9
<p>Field survey photos of ground subsidence in Zhuhai. (<b>a</b>,<b>b</b>) A water pump house in Pingsha Town, Jinwan District, Zhuhai City; (<b>c</b>,<b>d</b>) an elliptical-shaped building in the Fourteenth Village of Tanzhou Town, Zhongshan City; (<b>e</b>,<b>f</b>) a stilted house in Ma’an Village, Nanlang Town, Zhongshan City.</p> Full article ">Figure 10
<p>Deformation zones in Guangdong Province were identified by the automatic deformation detection method. The red points denote the location of deformation areas, and the blue lines are the administrative boundary.</p> Full article ">Figure 11
<p>Identified deformation zones in (<b>a</b>) Zhanjiang City, (<b>b</b>) Zhuhai City, and (<b>c</b>) Jiangmen City.</p> Full article ">
19 pages, 14195 KiB  
Article
A Transparent Soil Experiment to Investigate the Influence of Arrangement and Connecting Beams on the Pile–Soil Interaction of Micropile Groups
by Ziyi Wang, Jinqing Jia and Lihua Zhang
Sensors 2024, 24(16), 5448; https://doi.org/10.3390/s24165448 - 22 Aug 2024
Viewed by 523
Abstract
The use of a micropile group is an effective method for small and medium-sized slope management. However, there is limited research on the pile–soil interaction mechanism of micropile groups. Based on transparent soil and PIV technology, a test platform for the lateral load [...] Read more.
The use of a micropile group is an effective method for small and medium-sized slope management. However, there is limited research on the pile–soil interaction mechanism of micropile groups. Based on transparent soil and PIV technology, a test platform for the lateral load testing of slopes was constructed, and eight groups of transparent soil slope model experiments were performed. The changes in soil pressure and pile top displacement at the top of the piles during lateral loading were obtained. We scanned and photographed the slope, and obtained the deformation characteristics of the soil interior based on particle image velocimetry. A three-dimensional reconstruction program was developed to generate the displacement isosurface behind the pile. The impacts of various arrangement patterns and connecting beams on the deformation attributes and pile–soil interaction mechanism were explored, and the pile–soil interaction model of group piles was summarized. The results show that the front piles in a staggered arrangement bore more lateral thrust, and the distribution of soil pressure on each row of piles was more uniform. The connecting beams enhanced the overall stiffness of the pile group, reduced pile displacement, facilitated coordinated deformation of the pile group, and enhanced the anti-sliding effect of the pile–soil composite structure. Full article
(This article belongs to the Section Optical Sensors)
►▼ Show Figures

Figure 1

Figure 1
<p>Physical model test: 1—camera; 2—loading device; 3—PLC electrical box; 4—load cell; 5—acrylic model box; 6—model pills; 7—transparent soil; 8—laser device; 9—image acquisition system.</p> Full article ">Figure 2
<p>Transparent soil production.</p> Full article ">Figure 3
<p>Layout diagram of FSR.</p> Full article ">Figure 4
<p>Displacement of the double-row piles in the row arrangement.</p> Full article ">Figure 5
<p>Displacement of the double-row piles in the staggered arrangement.</p> Full article ">Figure 6
<p>Displacement of the three-row piles in the row arrangement.</p> Full article ">Figure 7
<p>Displacement of the three-row piles in the staggered arrangement.</p> Full article ">Figure 8
<p>Soil pressure of the double-row-arrangement piles.</p> Full article ">Figure 9
<p>Soil pressure of the piles with a double-row staggered arrangement.</p> Full article ">Figure 10
<p>Soil pressure of the three-row pile arrangement.</p> Full article ">Figure 11
<p>Soil pressure of the three-row staggered-arrangement piles.</p> Full article ">Figure 12
<p>Displacement cloud images of soil of the two-row pile group.</p> Full article ">Figure 13
<p>Displacement cloud images of soil in the three-row pile group.</p> Full article ">Figure 14
<p>Shear strain of the two-row pile group.</p> Full article ">Figure 15
<p>Shear strain of the three-row pile group.</p> Full article ">Figure 16
<p>Three-dimensional reconstruction of soil isosurfaces behind two rows of piles.</p> Full article ">Figure 17
<p>Three-dimensional reconstruction of soil isosurfaces behind the three-row piles.</p> Full article ">Figure 18
<p>Pile–soil interaction model of the micropile group.</p> Full article ">
23 pages, 17934 KiB  
Article
ACSwinNet: A Deep Learning-Based Rigid Registration Method for Head-Neck CT-CBCT Images in Image-Guided Radiotherapy
by Kuankuan Peng, Danyu Zhou, Kaiwen Sun, Junfeng Wang, Jianchun Deng and Shihua Gong
Sensors 2024, 24(16), 5447; https://doi.org/10.3390/s24165447 - 22 Aug 2024
Viewed by 474
Abstract
Accurate and precise rigid registration between head-neck computed tomography (CT) and cone-beam computed tomography (CBCT) images is crucial for correcting setup errors in image-guided radiotherapy (IGRT) for head and neck tumors. However, conventional registration methods that treat the head and neck as a [...] Read more.
Accurate and precise rigid registration between head-neck computed tomography (CT) and cone-beam computed tomography (CBCT) images is crucial for correcting setup errors in image-guided radiotherapy (IGRT) for head and neck tumors. However, conventional registration methods that treat the head and neck as a single entity may not achieve the necessary accuracy for the head region, which is particularly sensitive to radiation in radiotherapy. We propose ACSwinNet, a deep learning-based method for head-neck CT-CBCT rigid registration, which aims to enhance the registration precision in the head region. Our approach integrates an anatomical constraint encoder with anatomical segmentations of tissues and organs to enhance the accuracy of rigid registration in the head region. We also employ a Swin Transformer-based network for registration in cases with large initial misalignment and a perceptual similarity metric network to address intensity discrepancies and artifacts between the CT and CBCT images. We validate the proposed method using a head-neck CT-CBCT dataset acquired from clinical patients. Compared with the conventional rigid method, our method exhibits lower target registration error (TRE) for landmarks in the head region (reduced from 2.14 ± 0.45 mm to 1.82 ± 0.39 mm), higher dice similarity coefficient (DSC) (increased from 0.743 ± 0.051 to 0.755 ± 0.053), and higher structural similarity index (increased from 0.854 ± 0.044 to 0.870 ± 0.043). Our proposed method effectively addresses the challenge of low registration accuracy in the head region, which has been a limitation of conventional methods. This demonstrates significant potential in improving the accuracy of IGRT for head and neck tumors. Full article
(This article belongs to the Section Biomedical Sensors)
►▼ Show Figures

Figure 1

Figure 1
<p>Image-guided radiotherapy system.</p> Full article ">Figure 2
<p>Boundaries of the neck.</p> Full article ">Figure 3
<p>The architecture of the proposed Swin Transformer-based neural network with anatomical constraint for head-neck CT-CBCT images (ACSwinNet). The network incorporates a registration network based on Swin Transformer, while also including local constraint terms <span class="html-italic">L<sub>dice</sub></span> and global constraint terms <span class="html-italic">L<sub>ae</sub></span>. (<b>a</b>) The similarity metric network is used to calculate the perceptual loss between the fixed image and the transformed image (more detailed description is provided in <a href="#sec2dot2-sensors-24-05447" class="html-sec">Section 2.2</a>). (<b>b</b>) The anatomical constraint encoder is used to encode the global information of the segmentation map. (<b>c</b>) The registration network SwinNet is used to generate the transformation matrix <span class="html-italic">A</span>. It should be noted that the similarity metric network and the anatomical constraint encoder have been pretrained through a denoising autoencoder network (a more detailed description is provided in <a href="#sec2dot2-sensors-24-05447" class="html-sec">Section 2.2</a>). During the training of the registration network SwinNet, the weights of these two components remain fixed. In the inference phase of the model, the similarity metric network, the anatomical constraint encoder, and the segmentation maps are no longer required.</p> Full article ">Figure 4
<p>The architecture of Swin Transformer-based registration network. The network consists of four stages and is followed by a multilayer perceptron (MLP) module to generate the transformation parameters. Each stage is constructed of a convolutional layer and a Swin Transformer block. The output denoted as A signifies the parameters of the transformation matrix.</p> Full article ">Figure 5
<p>Swin Transformer block structure. (<b>a</b>) The self-attention mechanism is implemented in the (S)W-MSA module. (<b>b</b>) W-MSA and SW-MSA modules are embedded in the two successive Transformer blocks, respectively.</p> Full article ">Figure 6
<p>The details of the encoder–decoder architecture of a 3D convolutional de-noising auto-encoder (DAE) network with skip connections in between. It is trained to learn the representation of the input by restoring the original image without noise. <span class="html-italic">MSE</span> is chosen as the reconstruction loss. Once the DAE network is trained, the encoder can be used to extract the deep representation of the input.</p> Full article ">Figure 7
<p>A transformed image and a fixed image are input into the perceptual similarity metric network. Multi-scale feature maps of the input images are extracted by the convolutional layers and are involved in the similarity evaluation. L1 represents the L1 loss.</p> Full article ">Figure 8
<p>The results are produced by the ACSwinNet under different weighting parameters. The horizontal axis represents the different selections for hyperparameter <span class="html-italic">λ<sub>dice</sub></span>, while the various colors of the lines in the graph indicate the different selections for another hyperparameter, <span class="html-italic">λ<sub>ae</sub></span>. The vertical axis denotes the DSC values under these parameter settings.</p> Full article ">Figure 9
<p>The visualization of the CT image, the original CBCT image, and examples of the original CBCT image after undergoing various rotational and translational data augmentations.</p> Full article ">Figure 10
<p>Two CNN-based rigid registration network architectures. (<b>a</b>) Images are concatenated before being fed into the network (VTN). (<b>b</b>) Image features are extracted separately and then concatenated (ConvNet).</p> Full article ">Figure 11
<p>Box plots of different registration algorithms on different metrics.</p> Full article ">Figure 12
<p>A visual comparison of image registration outcomes in a typical case of significant initial misalignment is presented. CT and CBCT images are color-coded in magenta and green, respectively, to aid visual discrimination, while the overlapped region appears gray. The results are displayed in transverse, sagittal, and coronal views. The first column represents the fusion image before registration, while the subsequent columns depict the fused images of outcomes produced by different techniques.</p> Full article ">Figure 13
<p>The bar plot illustrates the average DSC of the anatomical structures located in the head region and neck region. Notably, the eye, optic nerve, and brainstem are located in the head region, while the spinal cord and larynx are located in the neck region.</p> Full article ">Figure 14
<p>A visual comparison of the image registration result of images with slight non-linear deformation in the neck area is presented. For better observation, CT and CBCT images are colored in magenta and green, respectively, while the overlapping region is shown in gray. Results are displayed in the view of the transverse, sagittal and coronal sections. The first column shows the fused image before registration, whereas the other columns show the fused images of results produced by different methods.</p> Full article ">Figure 15
<p>Slice examples of CT and CBCT feature maps at different scales. (<b>a</b>,<b>b</b>) The original image slices of CT and CBCT, respectively. (<b>c</b>–<b>h</b>) The feature maps at different resolution levels produced by the first three layers, with the resolution halved after each layer. Specifically, (<b>c</b>,<b>e</b>,<b>g</b>) are feature maps of CT images, while (<b>d</b>,<b>f</b>,<b>h</b>) are feature maps of CBCT images.</p> Full article ">Figure 15 Cont.
<p>Slice examples of CT and CBCT feature maps at different scales. (<b>a</b>,<b>b</b>) The original image slices of CT and CBCT, respectively. (<b>c</b>–<b>h</b>) The feature maps at different resolution levels produced by the first three layers, with the resolution halved after each layer. Specifically, (<b>c</b>,<b>e</b>,<b>g</b>) are feature maps of CT images, while (<b>d</b>,<b>f</b>,<b>h</b>) are feature maps of CBCT images.</p> Full article ">
39 pages, 2593 KiB  
Review
From Near-Sensor to In-Sensor: A State-of-the-Art Review of Embedded AI Vision Systems
by William Fabre, Karim Haroun, Vincent Lorrain, Maria Lepecq and Gilles Sicard
Sensors 2024, 24(16), 5446; https://doi.org/10.3390/s24165446 - 22 Aug 2024
Viewed by 1186
Abstract
In modern cyber-physical systems, the integration of AI into vision pipelines is now a standard practice for applications ranging from autonomous vehicles to mobile devices. Traditional AI integration often relies on cloud-based processing, which faces challenges such as data access bottlenecks, increased latency, [...] Read more.
In modern cyber-physical systems, the integration of AI into vision pipelines is now a standard practice for applications ranging from autonomous vehicles to mobile devices. Traditional AI integration often relies on cloud-based processing, which faces challenges such as data access bottlenecks, increased latency, and high power consumption. This article reviews embedded AI vision systems, examining the diverse landscape of near-sensor and in-sensor processing architectures that incorporate convolutional neural networks. We begin with a comprehensive analysis of the critical characteristics and metrics that define the performance of AI-integrated vision systems. These include sensor resolution, frame rate, data bandwidth, computational throughput, latency, power efficiency, and overall system scalability. Understanding these metrics provides a foundation for evaluating how different embedded processing architectures impact the entire vision pipeline, from image capture to AI inference. Our analysis delves into near-sensor systems that leverage dedicated hardware accelerators and commercially available components to efficiently process data close to their source, minimizing data transfer overhead and latency. These systems offer a balance between flexibility and performance, allowing for real-time processing in constrained environments. In addition, we explore in-sensor processing solutions that integrate computational capabilities directly into the sensor. This approach addresses the rigorous demand constraints of embedded applications by significantly reducing data movement and power consumption while also enabling in-sensor feature extraction, pre-processing, and CNN inference. By comparing these approaches, we identify trade-offs related to flexibility, power consumption, and computational performance. Ultimately, this article provides insights into the evolving landscape of embedded AI vision systems and suggests new research directions for the development of next-generation machine vision systems. Full article
(This article belongs to the Special Issue Sensor Technology for Intelligent Control and Computer Visions)
►▼ Show Figures

Figure 1

Figure 1
<p>Functional view of an integrated vision system with AI processing such as [<a href="#B23-sensors-24-05446" class="html-bibr">23</a>,<a href="#B24-sensors-24-05446" class="html-bibr">24</a>,<a href="#B25-sensors-24-05446" class="html-bibr">25</a>].</p> Full article ">Figure 2
<p>Design space for AI vision systems, highlighting the trade-off between flexibility and power consumption across processing configurations [<a href="#B23-sensors-24-05446" class="html-bibr">23</a>,<a href="#B29-sensors-24-05446" class="html-bibr">29</a>,<a href="#B31-sensors-24-05446" class="html-bibr">31</a>,<a href="#B32-sensors-24-05446" class="html-bibr">32</a>,<a href="#B33-sensors-24-05446" class="html-bibr">33</a>,<a href="#B34-sensors-24-05446" class="html-bibr">34</a>,<a href="#B35-sensors-24-05446" class="html-bibr">35</a>,<a href="#B36-sensors-24-05446" class="html-bibr">36</a>,<a href="#B37-sensors-24-05446" class="html-bibr">37</a>,<a href="#B38-sensors-24-05446" class="html-bibr">38</a>,<a href="#B39-sensors-24-05446" class="html-bibr">39</a>,<a href="#B40-sensors-24-05446" class="html-bibr">40</a>,<a href="#B41-sensors-24-05446" class="html-bibr">41</a>,<a href="#B42-sensors-24-05446" class="html-bibr">42</a>,<a href="#B43-sensors-24-05446" class="html-bibr">43</a>,<a href="#B44-sensors-24-05446" class="html-bibr">44</a>,<a href="#B45-sensors-24-05446" class="html-bibr">45</a>,<a href="#B46-sensors-24-05446" class="html-bibr">46</a>,<a href="#B47-sensors-24-05446" class="html-bibr">47</a>,<a href="#B48-sensors-24-05446" class="html-bibr">48</a>,<a href="#B49-sensors-24-05446" class="html-bibr">49</a>,<a href="#B50-sensors-24-05446" class="html-bibr">50</a>,<a href="#B51-sensors-24-05446" class="html-bibr">51</a>].</p> Full article ">Figure 3
<p>A basic CNN architecture. (<b>a</b>) Shows the network’s structure from input to output, visualizing data processing for predictions. (<b>b</b>) Details the mathematical operations, including weights, biases, and activation functions, that map inputs into outputs.</p> Full article ">Figure 4
<p>Representation of CNN models for embedded systems, showcasing memory requirements, computational demands, and accuracy [<a href="#B14-sensors-24-05446" class="html-bibr">14</a>,<a href="#B58-sensors-24-05446" class="html-bibr">58</a>,<a href="#B59-sensors-24-05446" class="html-bibr">59</a>,<a href="#B76-sensors-24-05446" class="html-bibr">76</a>,<a href="#B80-sensors-24-05446" class="html-bibr">80</a>,<a href="#B81-sensors-24-05446" class="html-bibr">81</a>,<a href="#B82-sensors-24-05446" class="html-bibr">82</a>,<a href="#B83-sensors-24-05446" class="html-bibr">83</a>,<a href="#B84-sensors-24-05446" class="html-bibr">84</a>,<a href="#B85-sensors-24-05446" class="html-bibr">85</a>,<a href="#B86-sensors-24-05446" class="html-bibr">86</a>] for classification tasks on ImageNet-1K [<a href="#B9-sensors-24-05446" class="html-bibr">9</a>]. Each bubble represents a model, with size indicating memory requirements and position reflecting ImageNet’s computational requirements and model accuracy (inspired by ([<a href="#B75-sensors-24-05446" class="html-bibr">75</a>]). Distinctive color patterns in <a href="#sensors-24-05446-f004" class="html-fig">Figure 4</a> categorize the models by their accuracy and implementation complexity, illustrating the trade-offs between accuracy, number of parameters, MAC requirements, and ease of layer implementation.</p> Full article ">Figure 5
<p>Representation of uniform and non-uniform quantization (inspired by [<a href="#B111-sensors-24-05446" class="html-bibr">111</a>]). The process converts full-precision weights and activations to lower bit-width representations, reducing the model size and computational requirements while preserving accuracy.</p> Full article ">Figure 6
<p>Representation of structured and unstructured pruning. Unstructured pruning removes individual weights across filters, allowing for fine-grained sparsity but potentially irregular computation patterns. Structured pruning removes entire channels or filters, resulting in a more regular pruned architecture that can be more efficiently implemented in hardware.</p> Full article ">Figure 7
<p>Simplified architectural view of near-sensor AI vision systems.</p> Full article ">Figure 8
<p>Simplified architectural view of in-sensor AI vision systems.</p> Full article ">
27 pages, 6870 KiB  
Article
Optimizing Indoor Airport Navigation with Advanced Visible Light Communication Systems
by Manuela Vieira, Manuel Augusto Vieira, Gonçalo Galvão, Paula Louro, Pedro Vieira and Alessandro Fantoni
Sensors 2024, 24(16), 5445; https://doi.org/10.3390/s24165445 - 22 Aug 2024
Viewed by 422
Abstract
This study presents a novel approach to enhancing indoor navigation in crowded multi-terminal airports using visible light communication (VLC) technology. By leveraging existing luminaires as transmission points, encoded messages are conveyed through modulated light signals to provide location-specific guidance. The objectives are to [...] Read more.
This study presents a novel approach to enhancing indoor navigation in crowded multi-terminal airports using visible light communication (VLC) technology. By leveraging existing luminaires as transmission points, encoded messages are conveyed through modulated light signals to provide location-specific guidance. The objectives are to facilitate navigation, optimize routes, and improve system performance through Edge/Fog integration. The methodology includes the use of tetrachromatic LED-equipped luminaires with On–Off Keying (OOK) modulation and a mesh cellular hybrid structure. Detailed airport modeling and user analysis (pedestrians and luggage/passenger carriers) equipped with PINPIN optical sensors are conducted. A VLC-specific communication protocol with coding and decoding techniques ensures reliable data transmission, while wayfinding algorithms offer real-time guidance. The results show effective data transmission and localization, enabling self-localization, travel direction inference, and route optimization. Agent-based simulations demonstrate improved traffic control, with analyses of user halting and average speed. This approach provides reliable indoor navigation independent of GPS signals, enhancing accessibility and convenience for airport users. The integration of VLC with Edge/Fog architecture ensures efficient movement through complex airport layouts. Full article
(This article belongs to the Special Issue Intelligent Sensors and Sensing Technologies in Vehicle Networks)
►▼ Show Figures

Figure 1

Figure 1
<p>(<b>a</b>) Three-dimensional relative positions between the transmitters and the receivers, footprints, and coverage map in the square topology. (<b>b</b>) Structure and operation of the PIN/PIN receiver. (<b>c</b>) Spectral gain under violet front optical bias (α<sup>V</sup>). The arrows point toward the optical gain at the analyzed R, G, B, and V input channels.</p> Full article ">Figure 2
<p>(<b>a</b>) Airport layout. (<b>b</b>) One-lane draft of the Edge/Frog hybrid architecture.</p> Full article ">Figure 3
<p>MUX/DEMUX signal of the calibrated cell. In the same frame of time, a random signal is superimposed [<a href="#B23-sensors-24-05445" class="html-bibr">23</a>].</p> Full article ">Figure 4
<p>(<b>a</b>) Simulated scenario and environment with the optical infrastructure (X<sub>ij</sub>), the generated footprints (1–9), and the connected luggage and pedestrian flow. (<b>b</b>) Terminal 2 schematic with coded lanes (L/0–7) and traffic lights (TL/0–15). Adapted from Ref. [<a href="#B28-sensors-24-05445" class="html-bibr">28</a>].</p> Full article ">Figure 5
<p>(<b>a</b>) Schematic of a one cycle phase diagram with eight moving carrier phases and an exclusive pedestrian phase. (<b>b</b>) Flowchart that illustrates the VLC wayfinding algorithm. (<b>c</b>) Flowchart to find the optimal path.</p> Full article ">Figure 6
<p>MUX signal request (<b>a</b>) and response (<b>b</b>) allocated to different V-VLC types. On the top, the decoded messages are displayed.</p> Full article ">Figure 7
<p>Normalized MUX signal responses and the corresponding decoded messages. (<b>a</b>) Messages sent (P<sub>1</sub>,<sub>2</sub> 2 I) by pedestrians waiting in the corners and (<b>b</b>) messages received by them (I 2 P<sub>1,2</sub>) at various frame times.</p> Full article ">Figure 8
<p>(<b>a</b>) SUMO environment. (<b>b</b>) State phasing diagram for the high-traffic scenario.</p> Full article ">Figure 9
<p>State phasing diagram for the low-traffic scenario.</p> Full article ">Figure 10
<p>High and low vehicular traffic scenario comparison as a function of the cycle’s duration. (<b>a</b>). Average pedestrian speed (11,200 p/h). (<b>b</b>) Average pedestrian speed (7200 p/h). (<b>c</b>) Halting (11,200 p/h). (<b>d</b>) Halting (7200 p/h).</p> Full article ">Figure 11
<p>High–high scenario (2300 carriers and 11,200 pedestrians per hour): (<b>a</b>) average speed of pedestrians and (<b>b</b>) density of pedestrians as a function of time in the target road and in the waiting corners. High–low scenario (2300 carriers and 7200 pedestrians per hour): (<b>c</b>) average speed of pedestrians and (<b>d</b>) density of pedestrians as a function of time in the target road and in the waiting corners.</p> Full article ">Figure 12
<p>High- and low-vehicular-traffic scenario comparison as a function of the cycle’s duration. (<b>a</b>) Halting (11,200 p/h). (<b>b</b>) Halting (7200 p/h).</p> Full article ">
30 pages, 13560 KiB  
Article
Beehive Smart Detector Device for the Detection of Critical Conditions That Utilize Edge Device Computations and Deep Learning Inferences
by Sotirios Kontogiannis
Sensors 2024, 24(16), 5444; https://doi.org/10.3390/s24165444 - 22 Aug 2024
Viewed by 612
Abstract
This paper presents a new edge detection process implemented in an embedded IoT device called Bee Smart Detection node to detect catastrophic apiary events. Such events include swarming, queen loss, and the detection of Colony Collapse Disorder (CCD) conditions. Two deep learning sub-processes [...] Read more.
This paper presents a new edge detection process implemented in an embedded IoT device called Bee Smart Detection node to detect catastrophic apiary events. Such events include swarming, queen loss, and the detection of Colony Collapse Disorder (CCD) conditions. Two deep learning sub-processes are used for this purpose. The first uses a fuzzy multi-layered neural network of variable depths called fuzzy-stranded-NN to detect CCD conditions based on temperature and humidity measurements inside the beehive. The second utilizes a deep learning CNN model to detect swarming and queen loss cases based on sound recordings. The proposed processes have been implemented into autonomous Bee Smart Detection IoT devices that transmit their measurements and the detection results to the cloud over Wi-Fi. The BeeSD devices have been tested for easy-to-use functionality, autonomous operation, deep learning model inference accuracy, and inference execution speeds. The author presents the experimental results of the fuzzy-stranded-NN model for detecting critical conditions and deep learning CNN models for detecting swarming and queen loss. From the presented experimental results, the stranded-NN achieved accuracy results up to 95%, while the ResNet-50 model presented accuracy results up to 99% for detecting swarming or queen loss events. The ResNet-18 model is also the fastest inference speed replacement of the ResNet-50 model, achieving up to 93% accuracy results. Finally, cross-comparison of the deep learning models with machine learning ones shows that deep learning models can provide at least 3–5% better accuracy results. Full article
(This article belongs to the Special Issue Deep-Learning-Based Defect Detection for Smart Manufacturing)
►▼ Show Figures

Figure 1

Figure 1
<p>Amplitude response over frequency inside a beehive in normal cases [<a href="#B59-sensors-24-05444" class="html-bibr">59</a>,<a href="#B60-sensors-24-05444" class="html-bibr">60</a>], of waggle dance [<a href="#B51-sensors-24-05444" class="html-bibr">51</a>,<a href="#B60-sensors-24-05444" class="html-bibr">60</a>,<a href="#B61-sensors-24-05444" class="html-bibr">61</a>], queen pipping tone [<a href="#B62-sensors-24-05444" class="html-bibr">62</a>,<a href="#B63-sensors-24-05444" class="html-bibr">63</a>], wings flapping under extreme temperature and humidity conditions [<a href="#B62-sensors-24-05444" class="html-bibr">62</a>], swarming [<a href="#B9-sensors-24-05444" class="html-bibr">9</a>,<a href="#B26-sensors-24-05444" class="html-bibr">26</a>,<a href="#B58-sensors-24-05444" class="html-bibr">58</a>,<a href="#B64-sensors-24-05444" class="html-bibr">64</a>], and due to poor nutrition [<a href="#B62-sensors-24-05444" class="html-bibr">62</a>].</p> Full article ">Figure 2
<p>Proposed Bee Smart Detection system’s high-level architecture. Bee Smart Detection system parts, data inputs, outputs, and data processing steps.</p> Full article ">Figure 3
<p>Proposed Bee Smart Detection device, device parts, and connected sensors; (<b>a</b>) BeeSD device design; and (<b>b</b>) BeeSD device prototype and its external battery source.</p> Full article ">Figure 4
<p>Illustrations of the BeeSD IoT device User Interfaces. These interfaces are presented to the beekeepers via the BeeSD system mobile phone application: (<b>a</b>) temperature, humidity, and sound intensity User Interface. (<b>b</b>) Mel spectrogram User Interface. (<b>c</b>) SFFT contributing frequencies’ mean amplitude values (called beegrams; mentioned in <a href="#sensors-24-05444-f001" class="html-fig">Figure 1</a>) User Interface. (<b>d</b>) Aggregated beegram correlation functions values over time corresponding to major indicators (growth, queen tone, hunger, thermal stress, total swarming) User Interfaces.</p> Full article ">Figure 5
<p>Bee Smart Detection processes for (1) queen loss and swarming (CNN model) and (2) stressful condition events (fuzzy-stranded-NN classifier).</p> Full article ">Figure 6
<p>Fuzzification process for data generation for the four fuzzy-stranded-NN classes: (<b>a</b>) temperature set, (<b>b</b>) humidity set, (<b>c</b>) purification process outputs (based on fuzzy rules) in three distinct humidity cases: low (40%), normal (60%), high (90%).</p> Full article ">Figure 7
<p>Bee Smart Detection prototypes tested in the beekeeping station of the laboratory team of MicroComputer Systems, Department of Mathematics, University of Ioannina, located in the Ligopsa area, Ioannina, Epirus, Greece.</p> Full article ">Figure 8
<p>Accuracy over trained epochs for CNN classification models detecting swarming and queen loss.</p> Full article ">Figure 9
<p>Classification loss over trained epochs for CNN classification models detecting swarming and queen loss.</p> Full article ">
17 pages, 3942 KiB  
Article
Study of Neutron-, Proton-, and Gamma-Irradiated Silicon Detectors Using the Two-Photon Absorption–Transient Current Technique
by Sebastian Pape, Marcos Fernández García, Michael Moll and Moritz Wiehe
Sensors 2024, 24(16), 5443; https://doi.org/10.3390/s24165443 - 22 Aug 2024
Viewed by 431
Abstract
The Two-Photon Absorption–Transient Current Technique (TPA-TCT) is a device characterisation technique that enables three-dimensional spatial resolution. Laser light in the quadratic absorption regime is employed to generate excess charge carriers only in a small volume around the focal spot. The drift of the [...] Read more.
The Two-Photon Absorption–Transient Current Technique (TPA-TCT) is a device characterisation technique that enables three-dimensional spatial resolution. Laser light in the quadratic absorption regime is employed to generate excess charge carriers only in a small volume around the focal spot. The drift of the excess charge carriers is studied to obtain information about the device under test. Neutron-, proton-, and gamma-irradiated p-type pad silicon detectors up to equivalent fluences of about 7 × 1015 neq/cm2 and a dose of 186 Mrad are investigated to study irradiation-induced effects on the TPA-TCT. Neutron and proton irradiation lead to additional linear absorption, which does not occur in gamma-irradiated detectors. The additional absorption is related to cluster damage, and the absorption scales according to the non-ionising energy loss. The influence of irradiation on the two-photon absorption coefficient is investigated, as well as potential laser beam depletion by the irradiation-induced linear absorption. Further, the electric field in neutron- and proton-irradiated pad detectors at an equivalent fluence of about 7 × 1015 neq/cm2 is investigated, where the space charge of the proton-irradiated devices appears inverted compared to the neutron-irradiated device. Full article
(This article belongs to the Special Issue Radiation Sensors and Detectors: Materials, Principles and Applications)
►▼ Show Figures

Figure 1

Figure 1
<p>Sketch of the used table-top TPA-TCT setup.</p> Full article ">Figure 2
<p>(<b>a</b>) Charge versus depth profile of an irradiated (3.32 × 10<sup>14</sup> <math display="inline"><semantics> <mrow> <mrow> <mi mathvariant="normal">n</mi> <mo>/</mo> <mi mathvariant="normal">c</mi> <mi mathvariant="normal">m</mi> </mrow> </mrow> </semantics></math>) <math display="inline"><semantics> <mrow> <mn>156</mn> <mo> </mo> <mrow> <mi mathvariant="sans-serif">μ</mi> <mi mathvariant="normal">m</mi> </mrow> </mrow> </semantics></math> thick p-type planar pad detector. The bias voltage is <math display="inline"><semantics> <mrow> <mn>300</mn> <mo> </mo> <mi mathvariant="normal">V</mi> </mrow> </semantics></math>. A comparison of the three different SPA correction methods is shown. The waveform subtraction method is indicated by the index WF, the subtraction method by the index const , and the correction by intensity method by the index I. The uncorrected data are given without index. (<b>b</b>) Time over threshold measurement in the same pad detector, including a comparison of the waveform subtraction and the intensity correction method.</p> Full article ">Figure 3
<p>Current transients recorded in the non-irradiated 21-DS-79 (<b>a</b>), the neutron-irradiated 21-DS-102, and the proton-irradiated 21-DS-92 (<b>b</b>) CiS pad detector. The fluence of the neutron- and proton-irradiated device were 7.02 × 10<sup>15</sup> <math display="inline"><semantics> <mrow> <mrow> <mi mathvariant="normal">n</mi> <mo>/</mo> <mi mathvariant="normal">c</mi> <msup> <mi mathvariant="normal">m</mi> <mn>2</mn> </msup> </mrow> </mrow> </semantics></math> and 1.17 × 10<sup>16</sup> <math display="inline"><semantics> <mrow> <mrow> <mi mathvariant="normal">p</mi> <mo>/</mo> <mi mathvariant="normal">c</mi> <msup> <mi mathvariant="normal">m</mi> <mn>2</mn> </msup> </mrow> </mrow> </semantics></math>, respectively. Positions in the legends refer to positions of the focal point where <math display="inline"><semantics> <mrow> <mn>0</mn> <mo> </mo> <mrow> <mi mathvariant="sans-serif">μ</mi> <mi mathvariant="normal">m</mi> </mrow> </mrow> </semantics></math> corresponds to the top side and <math display="inline"><semantics> <mrow> <mn>156</mn> <mo> </mo> <mrow> <mi mathvariant="sans-serif">μ</mi> <mi mathvariant="normal">m</mi> </mrow> </mrow> </semantics></math> to the back side of the sensor. The measurements were performed at <math display="inline"><semantics> <mrow> <mo>−</mo> <mn>20</mn> <mrow> <mo> </mo> <mo>°</mo> <mi mathvariant="normal">C</mi> </mrow> </mrow> </semantics></math> and 0% relative humidity. The beam parameters were <math display="inline"><semantics> <mrow> <msub> <mi>w</mi> <mn>0</mn> </msub> <mo>=</mo> <mn>1.2</mn> <mo> </mo> <mrow> <mi mathvariant="sans-serif">μ</mi> <mi mathvariant="normal">m</mi> </mrow> </mrow> </semantics></math> and <math display="inline"><semantics> <mrow> <msub> <mi>z</mi> <mi mathvariant="normal">R</mi> </msub> <mo>=</mo> <mn>9.7</mn> <mo> </mo> <mrow> <mi mathvariant="sans-serif">μ</mi> <mi mathvariant="normal">m</mi> </mrow> </mrow> </semantics></math>, and a pulse energy of <math display="inline"><semantics> <mrow> <mn>200</mn> <mo> </mo> <mrow> <mi mathvariant="normal">p</mi> <mi mathvariant="normal">J</mi> </mrow> </mrow> </semantics></math> was used. The laser frequency was <math display="inline"><semantics> <mrow> <mn>200</mn> <mo> </mo> <mi>Hz</mi> </mrow> </semantics></math>, and the average of 256 single acquisitions was recorded. The bias voltage was <math display="inline"><semantics> <mrow> <mn>300</mn> <mo> </mo> <mi mathvariant="normal">V</mi> </mrow> </semantics></math>. The signals are shifted on the time axis for better readability.</p> Full article ">Figure 4
<p>In-depth scans of the charge collection in pad detectors for neutron- (<b>a</b>) and proton- (<b>b</b>) irradiated samples. Figure (<b>c</b>,<b>d</b>) show the SPA corrected in-depth scans for neutrons and protons, respectively. The bias voltage is <math display="inline"><semantics> <mrow> <mn>300</mn> <mo> </mo> <mi mathvariant="normal">V</mi> </mrow> </semantics></math>.</p> Full article ">Figure 5
<p>Charge collected by SPA in neutron- (<b>a</b>) and proton- (<b>b</b>) irradiated <math display="inline"><semantics> <mrow> <mn>156</mn> <mo> </mo> <mrow> <mi mathvariant="sans-serif">μ</mi> <mi mathvariant="normal">m</mi> </mrow> </mrow> </semantics></math> thick pad sensors. Charge collected by TPA for the same neutron- (<b>c</b>) and proton- (<b>d</b>) irradiated devices. The TPA charge is extracted as the mean of the collected charge between the FWHM of the in-depth scans. The bias voltage for all scans is <math display="inline"><semantics> <mrow> <mn>300</mn> <mo> </mo> <mi mathvariant="normal">V</mi> </mrow> </semantics></math>.</p> Full article ">Figure 6
<p>Prompt current at <math display="inline"><semantics> <mrow> <msub> <mi>t</mi> <mi>pc</mi> </msub> <mo>=</mo> <mn>600</mn> <mo> </mo> <mrow> <mi mathvariant="normal">p</mi> <mi mathvariant="normal">s</mi> </mrow> </mrow> </semantics></math> for different bias voltages measured for (<b>a</b>) a neutron fluence of 7.02 × 10<sup>15</sup> <math display="inline"><semantics> <mrow> <mrow> <mi mathvariant="normal">n</mi> <mo>/</mo> <mi mathvariant="normal">c</mi> <msup> <mi mathvariant="normal">m</mi> <mn>2</mn> </msup> </mrow> </mrow> </semantics></math> and (<b>b</b>) a proton fluence of 1.17 × 10<sup>16</sup> <math display="inline"><semantics> <mrow> <mrow> <mi mathvariant="normal">p</mi> <mo>/</mo> <mi mathvariant="normal">c</mi> <msup> <mi mathvariant="normal">m</mi> <mn>2</mn> </msup> </mrow> </mrow> </semantics></math>. The equivalent fluences are comparable. The double junction is clearly visible in both plots. The proton-irradiated sample shows an electric field that grows from, and is stronger at, the back electrode. This effect is often called type inversion .</p> Full article ">Figure 7
<p>(<b>a</b>) In-depth scans of the charge collection in pad detectors. Gamma-irradiated samples and a non-irradiated sample are shown. (<b>b</b>) Comparison between the in-depth scans of a non-irradiated, a neutron-, a proton-, and a gamma-irradiated pad sensor. The bias voltage in all scans is <math display="inline"><semantics> <mrow> <mn>300</mn> <mo> </mo> <mi mathvariant="normal">V</mi> </mrow> </semantics></math>.</p> Full article ">Figure 8
<p>(<b>a</b>) In-depth scans of the charge collection in pad detectors. A non-irradiated sample is shown as well as a neutron-, a proton-, and a gamma-irradiated sample. (<b>b</b>) Same in-depth scans, but the SPA offset is corrected by the waveform subtraction method. The bias voltage is <math display="inline"><semantics> <mrow> <mn>300</mn> <mo> </mo> <mi mathvariant="normal">V</mi> </mrow> </semantics></math>.</p> Full article ">Figure 9
<p>(<b>a</b>) SPA charge collection normalised with the average electric field versus the equivalent fluence for neutron- and proton-irradiated FZ p-type pad detector. (<b>b</b>) TPA charge collection versus the equivalent fluence and dose for the same DUT.</p> Full article ">Figure 10
<p>Charge collected by SPA normalised with the pulse energy for proton and neutron irradiation in <math display="inline"><semantics> <mrow> <mn>156</mn> <mo> </mo> <mrow> <mi mathvariant="sans-serif">μ</mi> <mi mathvariant="normal">m</mi> </mrow> </mrow> </semantics></math> thick pad detectors, biased to <math display="inline"><semantics> <mrow> <mn>300</mn> <mo> </mo> <mi mathvariant="normal">V</mi> </mrow> </semantics></math>. The fit function is of the form <math display="inline"><semantics> <mrow> <mi>C</mi> <mspace width="0.166667em"/> <mo>·</mo> <mspace width="0.166667em"/> <msubsup> <mi mathvariant="sans-serif">Φ</mi> <mrow> <mi>eq</mi> </mrow> <mi mathvariant="normal">a</mi> </msubsup> </mrow> </semantics></math>, with <math display="inline"><semantics> <mrow> <mi>C</mi> <mo>=</mo> <mn>14.3</mn> <mo>×</mo> <msup> <mn>10</mn> <mrow> <mo>−</mo> <mn>11</mn> </mrow> </msup> <mo> </mo> <mrow> <mi mathvariant="normal">c</mi> <msup> <mi mathvariant="normal">m</mi> <mrow> <mn>2</mn> <mi mathvariant="normal">a</mi> </mrow> </msup> <mi mathvariant="normal">f</mi> <mi mathvariant="normal">C</mi> <mo>/</mo> <mi mathvariant="normal">n</mi> <mi mathvariant="normal">J</mi> </mrow> </mrow> </semantics></math> and <math display="inline"><semantics> <mrow> <mi>a</mi> <mo>=</mo> <mrow> <mn>0.84</mn> </mrow> </mrow> </semantics></math>. The highest fluences are excluded from the fit.</p> Full article ">Figure 11
<p>(<b>a</b>) Effective linear absorption coefficient of neutron-irradiated pad detectors versus the bias voltage, which is normalised with the device thickness. (<b>b</b>) Effective linear absorption coefficient for <math display="inline"><semantics> <mrow> <mn>156</mn> <mo> </mo> <mrow> <mi mathvariant="sans-serif">μ</mi> <mi mathvariant="normal">m</mi> </mrow> </mrow> </semantics></math> thick neutron- and proton-irradiated pad detectors versus the equivalent fluence. The value for the maximum bias voltage is used. The absorption coefficient does not saturate for the highest fluences, and the highest applied bias voltage is different for the proton- and neutron-irradiated devices. In order to show comparable <math display="inline"><semantics> <msub> <mi>α</mi> <mi>eff</mi> </msub> </semantics></math> for the highest fluences, a bias voltage of <math display="inline"><semantics> <mrow> <mn>300</mn> <mo> </mo> <mi mathvariant="normal">V</mi> </mrow> </semantics></math> is selected. The arrows are used to guide the eye towards the empty markers. These show the <math display="inline"><semantics> <msub> <mi>α</mi> <mi>eff</mi> </msub> </semantics></math> calculated from the interpolated <math display="inline"><semantics> <mrow> <msub> <mi>Q</mi> <mi>SPA</mi> </msub> <mrow> <mo>(</mo> <msub> <mi mathvariant="sans-serif">Φ</mi> <mi>eq</mi> </msub> <mo>)</mo> </mrow> </mrow> </semantics></math> of <a href="#sensors-24-05443-f010" class="html-fig">Figure 10</a>.</p> Full article ">Figure 12
<p>(<b>a</b>) Charge collection of a non-irradiated <math display="inline"><semantics> <mrow> <mn>156</mn> <mo> </mo> <mrow> <mi mathvariant="sans-serif">μ</mi> <mi mathvariant="normal">m</mi> </mrow> </mrow> </semantics></math> thick p-type pad sensor measured in a <sup>90</sup>Sr setup. (<b>b</b>) Charge collection in neutron-irradiated pad sensors for different fluences. The histogram is normalised to the maximum amount of counts to ease the comparison. The bias voltage in all measurements is <math display="inline"><semantics> <mrow> <mn>300</mn> <mo> </mo> <mi mathvariant="normal">V</mi> </mrow> </semantics></math>.</p> Full article ">Figure 13
<p>Charge collection efficiency for pad sensors, measured with the TPA-TCT and a <sup>90</sup>Sr setup, against the equivalent fluence. The type of irradiation is mentioned in the legend.</p> Full article ">Figure 14
<p>(<b>a</b>) Time over threshold profiles of the non-irradiated and the irradiated pad detectors. The highest fluences, while avoiding the double junction effect, are used to allow the comparison. The refraction from the air–silicon interface is not corrected in the <span class="html-italic">z</span>-axis. The location of the top and back interface is extracted from the non-irradiated device, and they are indicated by the dashed lines. (<b>b</b>) Refractive index extracted from the in-depth scans of various irradiated pad detectors. The nominal refractive index <math display="inline"><semantics> <mrow> <mi>n</mi> <mo>(</mo> <mn>250</mn> <mo> </mo> <mi mathvariant="normal">K</mi> <mo>)</mo> <mo>=</mo> <mrow> <mn>3.4681</mn> <mo>±</mo> <mn>0.0002</mn> </mrow> </mrow> </semantics></math> is taken from [<a href="#B31-sensors-24-05443" class="html-bibr">31</a>]. All scans are performed with a bias voltage of <math display="inline"><semantics> <mrow> <mn>300</mn> <mo> </mo> <mi mathvariant="normal">V</mi> </mrow> </semantics></math>.</p> Full article ">
28 pages, 8377 KiB  
Review
Research Progress on Saccharide Molecule Detection Based on Nanopores
by Bohua Yin, Wanyi Xie, Shaoxi Fang, Shixuan He, Wenhao Ma, Liyuan Liang, Yajie Yin, Daming Zhou, Zuobin Wang and Deqiang Wang
Sensors 2024, 24(16), 5442; https://doi.org/10.3390/s24165442 - 22 Aug 2024
Viewed by 677
Abstract
Saccharides, being one of the fundamental molecules of life, play essential roles in the physiological and pathological functions of cells. However, their intricate structures pose challenges for detection. Nanopore technology, with its high sensitivity and capability for single-molecule-level analysis, has revolutionized the identification [...] Read more.
Saccharides, being one of the fundamental molecules of life, play essential roles in the physiological and pathological functions of cells. However, their intricate structures pose challenges for detection. Nanopore technology, with its high sensitivity and capability for single-molecule-level analysis, has revolutionized the identification and structural analysis of saccharide molecules. This review focuses on recent advancements in nanopore technology for carbohydrate detection, presenting an array of methods that leverage the molecular complexity of saccharides. Biological nanopore techniques utilize specific protein binding or pore modifications to trigger typical resistive pulses, enabling the high-sensitivity detection of monosaccharides and oligosaccharides. In solid-state nanopore sensing, boronic acid modification and pH gating mechanisms are employed for the specific recognition and quantitative analysis of polysaccharides. The integration of artificial intelligence algorithms can further enhance the accuracy and reliability of analyses. Serving as a crucial tool in carbohydrate detection, we foresee significant potential in the application of nanopore technology for the detection of carbohydrate molecules in disease diagnosis, drug screening, and biosensing, fostering innovative progress in related research domains. Full article
(This article belongs to the Special Issue Micro- and Nano-Technologies for Sensing: From Device Fabrication to Applications)
►▼ Show Figures

Figure 1

Figure 1
<p>The roadmap presents the vital developmental milestones of nanopore sensing technology. Emerging as single-molecule sensors, nanopores have persistently enhanced their detection sensitivity and range of applications, aiming to realize the sequencing of fundamental biological components like nucleic acids, proteins, and polysaccharides.</p> Full article ">Figure 2
<p>Outlines the potential of nanopore technology for detecting various saccharide molecules. Mainly detecting four types of saccharide molecules including monosaccharides, oligosaccharides, polysaccharides, and glycosylated molecules. Nanopore technology mainly includes two detection techniques: biological nanopores and solid-state nanopores. These two nanopore detection techniques have made certain contributions to the detection of these four types of saccharide molecules.</p> Full article ">Figure 3
<p>Schematic of monosaccharide detection strategy based on nanopore technology. (<b>a</b>) Biological nanopore detection techniques rely on resistive pulse sensing; (<b>b</b>) solid-state nanopore detection techniques rely on ion rectification detection strategy (red line is the original rectification behavior; blue line is the charge changed line; green line is the structure behavior behavior.).</p> Full article ">Figure 4
<p>The detection of monosaccharides utilizing biological nanopores is depicted. (<b>a</b>) A direct quantitative detection method for glucose in body fluids was developed through the identification of protein conformation by ClyA nanopores and GBP−based nanopore sensing [<a href="#B46-sensors-24-05442" class="html-bibr">46</a>]. (<b>b</b>) Monosaccharide sensors were established by exploiting boronic acid-modified MspA nanopores and reversible covalent interactions between boronic acid and diols under aqueous conditions [<a href="#B51-sensors-24-05442" class="html-bibr">51</a>].</p> Full article ">Figure 5
<p>The detection of monosaccharides utilizing solid-state nanopores is depicted. (<b>a</b>) pH-gated glucose-responsive biomimetic nanopores engineered using chemically modified nanopores with APBA for glucose detection [<a href="#B56-sensors-24-05442" class="html-bibr">56</a>]. (<b>b</b>) The P(DMAEMA-<span class="html-italic">co</span>-VPBA) copolymer-modified biomimetic nanopores demonstrate rectification responses to pH, temperature, and the presence of sugar molecules [<a href="#B59-sensors-24-05442" class="html-bibr">59</a>]. (<b>c</b>) By utilizing a polymer modified with Au-S-bonded PATPBA-<span class="html-italic">co</span>-PNIPAAm, this nanopore sensor can detect glucose in saliva, prompting alterations in the conformation and wettability of the copolymer film to produce rectification responses [<a href="#B61-sensors-24-05442" class="html-bibr">61</a>].</p> Full article ">Figure 6
<p>The identification of oligosaccharides using nanopores is illustrated. (<b>a</b>) A spatial sensitivity evaluation of diverse isomeric saccharides binding conformations of MBP utilizing ClyA nanopores [<a href="#B63-sensors-24-05442" class="html-bibr">63</a>]. (<b>b</b>) Employing wild−type Ael nanopores for the direct discrimination of oligosaccharide derivatives subjected to labeled modifications [<a href="#B65-sensors-24-05442" class="html-bibr">65</a>]. (<b>c</b>) An examination of the distinctive absorption and conveyance of cyclic sugars facilitated by the CymA nanopore [<a href="#B66-sensors-24-05442" class="html-bibr">66</a>]. (<b>d</b>) An investigation into the pore attributes of glycosaminoglycans hexose, octaose, and decaose employing the AeL protein pore [<a href="#B69-sensors-24-05442" class="html-bibr">69</a>]. (<b>e</b>) Employing the porous structure of CN to examine two diagnostic trisaccharide epitope derivatives of LeA<sub>pN</sub> and SLeC<sub>pN</sub> [<a href="#B72-sensors-24-05442" class="html-bibr">72</a>].</p> Full article ">Figure 7
<p>The detection of heparin-like polysaccharides using solid-state nanopores. (<b>a</b>) Variations in event features’ distribution between heparin and OSCS samples. A solid-state silicon nitride nanopore sensor was employed for direct analysis of heparin and other polysaccharides using a straightforward statistical threshold algorithm [<a href="#B75-sensors-24-05442" class="html-bibr">75</a>]. (<b>b</b>) The utilization of a solid-state nanopore single-molecule sensor and SVM to develop a machine learning algorithm for the precise quantification and identification of GAG [<a href="#B76-sensors-24-05442" class="html-bibr">76</a>]. (<b>c</b>) Polysaccharide nanopore event data, portrayed as a scatter plot, were transformed into feature vectors, subsequently compressed using PCA for effective classification and recognition [<a href="#B78-sensors-24-05442" class="html-bibr">78</a>].</p> Full article ">Figure 8
<p>The detection of hyaluronic acid using solid-state nanopores. (<b>a</b>) A schematic diagram illustrating HA translocation through solid-state nanopores (left), alongside an ECD histogram for each molecular weight sample recorded at 200 mV, with colors corresponding to the molecular weight labels used in the gel image (right) [<a href="#B84-sensors-24-05442" class="html-bibr">84</a>]. (<b>b</b>) A schematic diagram showing the translocation of hyaluronic acid through glass conical nanopores modified with hyaluronidase on the tip side, along with an illustration of current blockage (top), with another schematic diagram displaying hyaluronic acid translocation through glass conical nanopores obstructed by hyaluronidase at the tip side, accompanied by an example of current blockade (bottom) [<a href="#B85-sensors-24-05442" class="html-bibr">85</a>].</p> Full article ">Figure 9
<p>Structure recognition of plant and other polysaccharides using solid-state nanopore. (<b>a</b>) Differentiation of xylitol molecule acetylation and deacetylation using solid-state nanopores [<a href="#B89-sensors-24-05442" class="html-bibr">89</a>]. (<b>b</b>) Investigation of detection of homogeneous and heterogeneous xylans utilizing glass nanopores modified with polyethyleneimine (PEI) [<a href="#B90-sensors-24-05442" class="html-bibr">90</a>]. (<b>c</b>) Biomimetic nanochannel system incorporating the responsive polymer Glc-PEI for accurately identifying sialic polysaccharides [<a href="#B94-sensors-24-05442" class="html-bibr">94</a>].</p> Full article ">Figure 10
<p>Glycosylation peptide detection based on biological nanopore. (<b>a</b>) The detection of glycosylated and phosphorylated peptides at the single-molecule level using the FraC biological nanopore from crustacean toxin is detailed [<a href="#B102-sensors-24-05442" class="html-bibr">102</a>]. Additionally, (<b>b</b>) phenylalanine-modified crustacean toxin FraC nanopores differentiate hydrophilic peptides from glycosylated peptides [<a href="#B103-sensors-24-05442" class="html-bibr">103</a>].</p> Full article ">Figure 11
<p>Schematic diagram of label-free detection of single-molecule glycoproteins based on boronic ester affinity interaction using thiol phenylboronic acid-modified glass conical nanopores, and current time curve of IgG in 4-MPBA-modified glass conical nanopores [<a href="#B104-sensors-24-05442" class="html-bibr">104</a>].</p> Full article ">
14 pages, 11962 KiB  
Article
Design and Validation of Single-Axis 3D-Printed Force Sensor Based on Three Nested Flexible Rings
by Pengfei Yang, Shiwei Xin, Yuqing Mao, Fei Dang and Feng Huang
Sensors 2024, 24(16), 5441; https://doi.org/10.3390/s24165441 - 22 Aug 2024
Viewed by 514
Abstract
Force measurement is crucial in numerous engineering applications, while traditional force sensors often face problems such as elevated expenses or significant measurement errors. To tackle this issue, we propose an innovative force sensor employing three nested flexible rings fabricated through 3D additive manufacturing, [...] Read more.
Force measurement is crucial in numerous engineering applications, while traditional force sensors often face problems such as elevated expenses or significant measurement errors. To tackle this issue, we propose an innovative force sensor employing three nested flexible rings fabricated through 3D additive manufacturing, which detects external forces through the displacement variations of flexible rings. An analytical model on the basis of the minimal energy method is developed to elucidate the force-displacement correlation with nonlinearity. Both FEM simulations and experiments verify the sensor’s effectiveness. This sensor has the advantages of low expenses and easy manufacture, indicating promising prospects in a range of applications, including robotics, the automotive industry, and iatrical equipment. Full article
(This article belongs to the Section Physical Sensors)
►▼ Show Figures

Figure 1

Figure 1
<p>Diagram for (<b>a</b>) FS and (<b>b</b>) flexible rings. Diagrams of three rings subjected to (<b>c</b>) the <span class="html-italic">F</span><sub>compression</sub> and (<b>d</b>) <span class="html-italic">F</span><sub>tensile</sub>.</p> Full article ">Figure 2
<p>(<b>a</b>) FEM Meshes of the three rings model. Displacement nephogram of three flexible rings under (<b>b</b>) <span class="html-italic">F</span><sub>compression</sub> of 20 N and (<b>c</b>) <span class="html-italic">F</span><sub>tensile</sub> of 20 N.</p> Full article ">Figure 3
<p>Force-displacement relationship comparison between simulation results and theoretical predictions.</p> Full article ">Figure 4
<p>(<b>a</b>) Experiment setup for <span class="html-italic">F</span><sub>compression</sub>. (<b>b</b>) Experiment setup for <span class="html-italic">F</span><sub>tensile</sub>. Scale bars: 30 mm.</p> Full article ">Figure 5
<p>Force-displacement relationship comparison between experimental results and theoretical predictions.</p> Full article ">Figure 6
<p>(<b>a</b>) Hysteresis test. (<b>b</b>) Repeatability test. (<b>c</b>) Temperature test. (42 °C and 20 °C).</p> Full article ">Figure 7
<p>(<b>a</b>) Cyclic test. (<b>b</b>) <span class="html-italic">F</span><sub>compression</sub> creep test. (<b>c</b>) <span class="html-italic">F</span><sub>tensile</sub> creep test.</p> Full article ">
20 pages, 3099 KiB  
Article
Diffusion Models-Based Purification for Common Corruptions on Robust 3D Object Detection
by Mumuxin Cai, Xupeng Wang, Ferdous Sohel and Hang Lei
Sensors 2024, 24(16), 5440; https://doi.org/10.3390/s24165440 - 22 Aug 2024
Viewed by 625
Abstract
LiDAR sensors have been shown to generate data with various common corruptions, which seriously affect their applications in 3D vision tasks, particularly object detection. At the same time, it has been demonstrated that traditional defense strategies, including adversarial training, are prone to suffering [...] Read more.
LiDAR sensors have been shown to generate data with various common corruptions, which seriously affect their applications in 3D vision tasks, particularly object detection. At the same time, it has been demonstrated that traditional defense strategies, including adversarial training, are prone to suffering from gradient confusion during training. Moreover, they can only improve their robustness against specific types of data corruption. In this work, we propose LiDARPure, which leverages the powerful generation ability of diffusion models to purify corruption in the LiDAR scene data. By dividing the entire scene into voxels to facilitate the processes of diffusion and reverse diffusion, LiDARPure overcomes challenges induced from adversarial training, such as sparse point clouds in large-scale LiDAR data and gradient confusion. In addition, we utilize the latent geometric features of a scene as a condition to assist the generation of diffusion models. Detailed experiments show that LiDARPure can effectively purify 19 common types of LiDAR data corruption. Further evaluation results demonstrate that it can improve the average precision of 3D object detectors to an extent of 20% in the face of data corruption, much higher than existing defence strategies. Full article
(This article belongs to the Special Issue Vision Sensors for Object Detection and Tracking)
►▼ Show Figures

Figure 1

Figure 1
<p>Overview of the proposed LiDARPure method for purifying the corrupted LiDAR scene. Given the corrupted LiDAR scene, it is divided into independent point cloud units through voxelization. LiDARPure utilizes a diffusion model to purify the point cloud within each unit. The corrupted points within the unit are converted into noise through a forward diffusion process, and then restored to purified point clouds through reverse diffusion process. The latent feature of the point cloud within the unit are extracted using the encoder and participate as the condition in the reverse diffusion process. Each unit and its latent features before (<math display="inline"><semantics> <msub> <mi>z</mi> <mn>0</mn> </msub> </semantics></math>) and after (<math display="inline"><semantics> <mover> <mi>z</mi> <mo>¯</mo> </mover> </semantics></math>) purification are used to calculate the denoising recovery loss. All units complete the diffusion process to form a purified scene.</p> Full article ">Figure 2
<p>The qualitative results of the proposed LiDARPure method purify the LiDAR scene corrupted by snowfall. The units used are indicated in the brackets.</p> Full article ">Figure 3
<p>The qualitative results of proposed LiDARPure method purifies the LiDAR scene corrupted by Gaussian noise. The units used are indicated in bracket.</p> Full article ">Figure 4
<p>Robust AP affected by different steps <span class="html-italic">T</span> on PV-RCNN detector in the reverse diffusion process.</p> Full article ">
22 pages, 13134 KiB  
Article
Syntax-Guided Content-Adaptive Transform for Image Compression
by Yunhui Shi, Liping Ye, Jin Wang, Lilong Wang, Hui Hu, Baocai Yin and Nam Ling
Sensors 2024, 24(16), 5439; https://doi.org/10.3390/s24165439 - 22 Aug 2024
Viewed by 474
Abstract
The surge in image data has significantly increased the pressure on storage and transmission, posing new challenges for image compression technology. The structural texture of an image implies its statistical characteristics, which is effective for image encoding and decoding. Consequently, content-adaptive compression methods [...] Read more.
The surge in image data has significantly increased the pressure on storage and transmission, posing new challenges for image compression technology. The structural texture of an image implies its statistical characteristics, which is effective for image encoding and decoding. Consequently, content-adaptive compression methods based on learning can better capture the content attributes of images, thereby enhancing encoding performance. However, learned image compression methods do not comprehensively account for both the global and local correlations among the pixels within an image. Moreover, they are constrained by rate-distortion optimization, which prevents the attainment of a compact representation of image attributes. To address these issues, we propose a syntax-guided content-adaptive transform framework that efficiently captures image attributes and enhances encoding efficiency. Firstly, we propose a syntax-refined side information module that fully leverages syntax and side information to guide the adaptive transformation of image attributes. Moreover, to more thoroughly exploit the global and local correlations in image space, we designed global–local modules, local–global modules, and upsampling/downsampling modules in codecs, further eliminating local and global redundancies. The experimental findings indicate that our proposed syntax-guided content-adaptive image compression model successfully adapts to the diverse complexities of different images, which enhances the efficiency of image compression. Concurrently, the method proposed has demonstrated outstanding performance across three benchmark datasets. Full article
(This article belongs to the Section Sensing and Imaging)
►▼ Show Figures

Figure 1

Figure 1
<p>The syntax-guided content-adaptive transform model architecture.</p> Full article ">Figure 2
<p>The syntax-refined side information module.</p> Full article ">Figure 3
<p>The G-LAM and L-GAM architecture.</p> Full article ">Figure 4
<p>The attention masks are computed in a local window.</p> Full article ">Figure 5
<p>The downsampling and upsampling module architecture.</p> Full article ">Figure 6
<p>The PSNR optimized RD curve on the Kodak dataset.</p> Full article ">Figure 7
<p>The PSNR optimized RD curve on the CLIC dataset.</p> Full article ">Figure 8
<p>The PSNR optimized RD curve on the Tecnick dataset.</p> Full article ">Figure 9
<p>The MS-SSIM optimized RD curve on the Kodak dataset.</p> Full article ">Figure 10
<p>The MS-SSIM optimized RD curve on the CLIC dataset.</p> Full article ">Figure 11
<p>The MS-SSIM optimized RD curve on the Tecnick dataset.</p> Full article ">Figure 12
<p>Subjective evaluation of Kodim24, which is optimized for PSNR. We compare our SGCATM with JPEG [<a href="#B14-sensors-24-05439" class="html-bibr">14</a>], BPG [<a href="#B16-sensors-24-05439" class="html-bibr">16</a>], WebP [<a href="#B17-sensors-24-05439" class="html-bibr">17</a>], VVC [<a href="#B18-sensors-24-05439" class="html-bibr">18</a>], Context-Adaptive Entropy [<a href="#B24-sensors-24-05439" class="html-bibr">24</a>], and Neural Syntax [<a href="#B41-sensors-24-05439" class="html-bibr">41</a>].</p> Full article ">Figure 13
<p>PSNR–bpp during testing on Kodim24.</p> Full article ">Figure 14
<p>Subjective evaluation of CLIC35, which is optimized for PSNR. We compare our SGCATM with JPEG [<a href="#B14-sensors-24-05439" class="html-bibr">14</a>], BPG [<a href="#B16-sensors-24-05439" class="html-bibr">16</a>], WebP [<a href="#B17-sensors-24-05439" class="html-bibr">17</a>], VVC [<a href="#B18-sensors-24-05439" class="html-bibr">18</a>], Coarse-to-Fine Hyper-Prior [<a href="#B26-sensors-24-05439" class="html-bibr">26</a>], and Neural Syntax [<a href="#B41-sensors-24-05439" class="html-bibr">41</a>].</p> Full article ">Figure 15
<p>PSNR–bpp during testing on CLIC35.</p> Full article ">Figure 16
<p>Subjective evaluation of Kodim7, which is optimized for MS-SSIM. We compare our SGCATM with JPEG [<a href="#B14-sensors-24-05439" class="html-bibr">14</a>], BPG [<a href="#B16-sensors-24-05439" class="html-bibr">16</a>], WebP [<a href="#B17-sensors-24-05439" class="html-bibr">17</a>], VVC [<a href="#B18-sensors-24-05439" class="html-bibr">18</a>], Context-Adaptive Entropy [<a href="#B24-sensors-24-05439" class="html-bibr">24</a>], and Neural Syntax [<a href="#B41-sensors-24-05439" class="html-bibr">41</a>].</p> Full article ">Figure 17
<p>MS-SSIM–bpp during testing on Kodim7.</p> Full article ">Figure 18
<p>Subjective evaluation of CLIC39, which is optimized for MS-SSIM. We compare our SGCATM with JPEG [<a href="#B14-sensors-24-05439" class="html-bibr">14</a>], BPG [<a href="#B16-sensors-24-05439" class="html-bibr">16</a>], WebP [<a href="#B17-sensors-24-05439" class="html-bibr">17</a>], VVC [<a href="#B18-sensors-24-05439" class="html-bibr">18</a>], Coarse-to-Fine Hyper-Prior [<a href="#B26-sensors-24-05439" class="html-bibr">26</a>], and Neural Syntax [<a href="#B41-sensors-24-05439" class="html-bibr">41</a>].</p> Full article ">Figure 19
<p>MS-SSIM–bpp during testing on CLIC39.</p> Full article ">Figure 20
<p>Ablation studies.</p> Full article ">Figure 21
<p>Bitrate allocation and visualization of reconstruction details of the proposed module for the channel with the latent of maximal entropy. The red frame signifies ’Reconstructed Details’.</p> Full article ">
16 pages, 5114 KiB  
Article
Comparative Analysis of Machine Learning and Deep Learning Algorithms for Assessing Agricultural Product Quality Using NIRS
by Jiwen Ren, Yuming Xiong, Xinyu Chen and Yong Hao
Sensors 2024, 24(16), 5438; https://doi.org/10.3390/s24165438 - 22 Aug 2024
Viewed by 734
Abstract
The success of near-infrared spectroscopy (NIRS) analysis hinges on the precision and robustness of the calibration model. Shallow learning (SL) algorithms like partial least squares discriminant analysis (PLS-DA) often fall short in capturing the interrelationships between adjacent spectral variables, and the analysis results [...] Read more.
The success of near-infrared spectroscopy (NIRS) analysis hinges on the precision and robustness of the calibration model. Shallow learning (SL) algorithms like partial least squares discriminant analysis (PLS-DA) often fall short in capturing the interrelationships between adjacent spectral variables, and the analysis results are easily affected by spectral noise, which dramatically limits the breadth and depth of applications of NIRS. Deep learning (DL) methods, with their capacity to discern intricate features from limited samples, have been progressively integrated into NIRS. In this paper, two discriminant analysis problems, including wheat kernels and Yali pears as examples, and several representative calibration models were used to research the robustness and effectiveness of the model. Additionally, this article proposed a near-infrared calibration model, which was based on the Gramian angular difference field method and coordinate attention convolutional neural networks (G-CACNNs). The research results show that, compared with SL, spectral preprocessing has a smaller impact on the analysis accuracy of consensus learning (CL) and DL, and the latter has the highest analysis accuracy in the modeling results using the original spectrum. The accuracy of G-CACNNs in two discrimination tasks was 98.48% and 99.39%. Finally, this research compared the performance of various models under noise to evaluate the robustness and noise resistance of the proposed method. Full article
(This article belongs to the Section Smart Agriculture)
►▼ Show Figures

Figure 1

Figure 1
<p>Network structure and CA internal structure. The CNN is composed of a series of convolutional layers and max pooling layers. The fully connected layer is then utilized for the final classification. By incorporating a CA module at the front of the network, the CACNNA network is formed. The CA module embeds positional information by pooling, concatenation, and convolution operations on the initial feature map in two directions.</p> Full article ">Figure 2
<p>The spectral collection system for Yali pears.</p> Full article ">Figure 3
<p>Experimental flow chart. The experiment is mainly composed of four stages. Initially, the spectra for agricultural products is collected. Following that, a variety of processing techniques are utilized to treat the spectra. Subsequently, discriminative models that are representative are developed according to the modeling strategy. Finally, an in-depth analysis of the modeling process is performed.</p> Full article ">Figure 4
<p>Average spectra. (<b>a</b>) The average spectra of wheat kernels. (<b>b</b>) The average spectra of Yali pears.</p> Full article ">Figure 5
<p>Schematic diagram of GAF converting process. The color change from blue to red corresponds to the increment of the value in the pixel.</p> Full article ">Figure 6
<p>Grad-CAM of DL model. The yellow area is paid attention to by the network, and the blue area network lacks attention. The red line is the dividing line of the attention heatmap. The attention concentration of the model is evaluated by dividing the heatmap into regions of interest and regions of non-interest.</p> Full article ">Figure 7
<p>Schematic diagram of additive noise of spectra. (<b>a</b>) Wheat kernel dataset. (<b>b</b>) Yali pear dataset.</p> Full article ">Figure 8
<p>Grad-CAM results with noise spectrum. The yellow area is paid attention to by the network, and the blue area network lacks attention. (<b>a</b>) G-CNN. (<b>b</b>) G-CACNN.</p> Full article ">Figure 9
<p>Results of different methods under different levels of noise. (<b>a</b>) Robustness test of the models based on the wheat kernel dataset. (<b>b</b>) Robustness test of the models based on the pear dataset.</p> Full article ">
21 pages, 726 KiB  
Systematic Review
Smart Spare Parts (SSP) in the Context of Industry 4.0: A Systematic Review
by G. Morales Pavez and Orlando Durán
Sensors 2024, 24(16), 5437; https://doi.org/10.3390/s24165437 - 22 Aug 2024
Viewed by 665
Abstract
The implementation of Industry 4.0 has integrated manufacturing, electronics, and engineering materials, leading to the creation of smart parts (SPs) that provide information on production system conditions. However, SP development faces challenges due to limitations in manufacturing processes and integrating electronic components. This [...] Read more.
The implementation of Industry 4.0 has integrated manufacturing, electronics, and engineering materials, leading to the creation of smart parts (SPs) that provide information on production system conditions. However, SP development faces challenges due to limitations in manufacturing processes and integrating electronic components. This systematic review synthesizes scientific articles on SP fabrication using additive manufacturing (AM), identifying the advantages and disadvantages of AM techniques in SP production and distinguishing between SPs and smart spare parts (SSPs). The methodology involves establishing a reference framework, formulating SP-related questions, and applying inclusion criteria and keywords, initially resulting in 1603 articles. After applying exclusion criteria, 70 articles remained. The results show that while SP development is advancing, widespread application of AM-manufactured SP is recent. SPs can anticipate production system failures, minimize design artifacts, and reduce manufacturing costs. Furthermore, the review highlights that SSPs, a subcategory of SPs, primarily differs by replacing conventional critical parts in the industry, offering enhanced functionality and reliability in industrial applications. The study concludes that continued research and development in this field is essential for further advancements and broader adoption of these technologies. Full article
(This article belongs to the Special Issue Sensing in Intelligent and Unmanned Additive Manufacturing)
►▼ Show Figures

Figure 1

Figure 1
<p>Venn diagram showing the convergence zone among different areas.</p> Full article ">Figure 2
<p>PRISMA flow diagram.</p> Full article ">Figure 3
<p>Stages for the introduction of a sensor inside the part.</p> Full article ">Figure 4
<p>SSPs are considered a subcategory of SPs.</p> Full article ">
34 pages, 2564 KiB  
Article
Achieving More with Less: A Lightweight Deep Learning Solution for Advanced Human Activity Recognition (HAR)
by Sarab AlMuhaideb, Lama AlAbdulkarim, Deemah Mohammed AlShahrani, Hessah AlDhubaib and Dalal Emad AlSadoun
Sensors 2024, 24(16), 5436; https://doi.org/10.3390/s24165436 - 22 Aug 2024
Viewed by 889
Abstract
Human activity recognition (HAR) is a crucial task in various applications, including healthcare, fitness, and the military. Deep learning models have revolutionized HAR, however, their computational complexity, particularly those involving BiLSTMs, poses significant challenges for deployment on resource-constrained devices like smartphones. While BiLSTMs [...] Read more.
Human activity recognition (HAR) is a crucial task in various applications, including healthcare, fitness, and the military. Deep learning models have revolutionized HAR, however, their computational complexity, particularly those involving BiLSTMs, poses significant challenges for deployment on resource-constrained devices like smartphones. While BiLSTMs effectively capture long-term dependencies by processing inputs bidirectionally, their high parameter count and computational demands hinder practical applications in real-time HAR. This study investigates the approximation of the computationally intensive BiLSTM component in a HAR model by using a combination of alternative model components and data flipping augmentation. The proposed modifications to an existing hybrid model architecture replace the BiLSTM with standard and residual LSTM, along with convolutional networks, supplemented by data flipping augmentation to replicate the context awareness typically provided by BiLSTM networks. The results demonstrate that the residual LSTM (ResLSTM) model achieves superior performance while maintaining a lower computational complexity compared to the traditional BiLSTM model. Specifically, on the UCI-HAR dataset, the ResLSTM model attains an accuracy of 96.34% with 576,702 parameters, outperforming the BiLSTM model’s accuracy of 95.22% with 849,534 parameters. On the WISDM dataset, the ResLSTM achieves an accuracy of 97.20% with 192,238 parameters, compared to the BiLSTM’s 97.23% accuracy with 283,182 parameters, demonstrating a more efficient architecture with minimal performance trade-off. For the KU-HAR dataset, the ResLSTM model achieves an accuracy of 97.05% with 386,038 parameters, showing comparable performance to the BiLSTM model’s 98.63% accuracy with 569,462 parameters, but with significantly fewer parameters. Full article
(This article belongs to the Special Issue Intelligent Wearable Sensor-Based Gait and Movement Analysis)
►▼ Show Figures

Figure 1

Figure 1
<p>Proposed model architecture for single sensor and data augmentation.</p> Full article ">Figure 2
<p>Proposed model architecture for multi-sensor and data augmentation.</p> Full article ">Figure 3
<p>Performance comparison of the different models in each group on the UCI-HAR dataset. (<b>a</b>) Accuracy scores. (<b>b</b>) Number of parameters.</p> Full article ">Figure 4
<p>Confusion matrix of the model in Group A as evaluated on the UCI-HAR dataset. (<b>a</b>) Using BiLSTM component. (<b>b</b>) Using ResBiLSTM component.</p> Full article ">Figure 5
<p>Confusion matrix of the model in Group B as evaluated on the UCI-HAR dataset. (<b>a</b>) Using LSTM component. (<b>b</b>) Using ResLSTM component.</p> Full article ">Figure 6
<p>Confusion matrix of the model in Group C as evaluated on the UCI-HAR dataset. (<b>a</b>) Using CNN component. (<b>b</b>) Using ResCNN component.</p> Full article ">Figure 7
<p>Performance comparison of the different models in each group using the WISDM dataset. (<b>a</b>) <math display="inline"><semantics> <mrow> <mi>F</mi> <mn>1</mn> </mrow> </semantics></math>-scores. (<b>b</b>) Number of parameters.</p> Full article ">Figure 8
<p>Confusion matrix of the model in Group A, as evaluated using the WISDM dataset. (<b>a</b>) Using BiLSTM component. (<b>b</b>) Using ResBiLSTM component.</p> Full article ">Figure 9
<p>Confusion matrix of the model in Group B as evaluated on the WISDM dataset. (<b>a</b>) Using LSTM component. (<b>b</b>) Using ResLSTM component.</p> Full article ">Figure 10
<p>Confusion matrix of the model in Group C as evaluated on the WISDM dataset. (<b>a</b>) Using CNN component. (<b>b</b>) Using ResCNN component.</p> Full article ">Figure 11
<p>Performance comparison of the different models in each group on the KU-HAR dataset. (<b>a</b>) <math display="inline"><semantics> <mrow> <mi>F</mi> <mn>1</mn> </mrow> </semantics></math>-scores. (<b>b</b>) Number of parameters.</p> Full article ">Figure 12
<p>Confusion matrix of the model in Group A as evaluated on the KU-HAR dataset. (<b>a</b>) Using BiLSTM component. (<b>b</b>) Using ResBiLSTM component.</p> Full article ">Figure 13
<p>Confusion matrix of the model in Group B as evaluated on the KU-HAR dataset. (<b>a</b>) Using LSTM component. (<b>b</b>) Using ResLSTM component.</p> Full article ">Figure 14
<p>Confusion matrix of the model in Group C as evaluated on the KU-HAR dataset. (<b>a</b>) Using CNN component. (<b>b</b>) Using ResCNN component.</p> Full article ">Figure 15
<p>Combined average metrics and parameters across all datasets. (<b>a</b>) Average performance metrics. (<b>b</b>) Number of parameters.</p> Full article ">Figure A1
<p>Training loss curves for the implemented models using UCI-HAR dataset. (<b>a</b>) BiLSTM, (<b>b</b>) ResBiLSTM, (<b>c</b>) LSTM, (<b>d</b>) ResLSTM, (<b>e</b>) CNN, (<b>f</b>) ResCNN.</p> Full article ">Figure A2
<p>Training accuracy curves for the implemented models using UCI-HAR dataset. (<b>a</b>) BiLSTM, (<b>b</b>) ResBiLSTM, (<b>c</b>) LSTM, (<b>d</b>) ResLSTM, (<b>e</b>) CNN, (<b>f</b>) ResCNN.</p> Full article ">Figure A3
<p>Training loss curves for the implemented models using WISDM dataset. (<b>a</b>) BiLSTM, (<b>b</b>) ResBiLSTM, (<b>c</b>) LSTM, (<b>d</b>) ResLSTM, (<b>e</b>) CNN, (<b>f</b>) ResCNN.</p> Full article ">Figure A4
<p>Training accuracy curves for the implemented models using WISDM dataset. (<b>a</b>) BiLSTM, (<b>b</b>) ResBiLSTM, (<b>c</b>) LSTM, (<b>d</b>) ResLSTM, (<b>e</b>) CNN, (<b>f</b>) ResCNN.</p> Full article ">Figure A5
<p>Training loss curves for the implemented models using KU-HAR dataset. (<b>a</b>) BiLSTM, (<b>b</b>) ResBiLSTM, (<b>c</b>) LSTM, (<b>d</b>) ResLSTM, (<b>e</b>) CNN, (<b>f</b>) ResCNN.</p> Full article ">Figure A6
<p>Training accuracy curves for the implemented models using KU-HAR dataset. (<b>a</b>) BiLSTM, (<b>b</b>) ResBiLSTM, (<b>c</b>) LSTM, (<b>d</b>) ResLSTM, (<b>e</b>) CNN, (<b>f</b>) ResCNN.</p> Full article ">
12 pages, 1677 KiB  
Article
Validity and Test–Retest Reliability of Spatiotemporal Running Parameter Measurement Using Embedded Inertial Measurement Unit Insoles
by Louis Riglet, Baptiste Orliac, Corentin Delphin, Audrey Leonard, Nicolas Eby, Paul Ornetti, Davy Laroche and Mathieu Gueugnon
Sensors 2024, 24(16), 5435; https://doi.org/10.3390/s24165435 - 22 Aug 2024
Viewed by 616
Abstract
Running is the basis of many sports and has highly beneficial effects on health. To increase the understanding of running, DSPro® insoles were developed to collect running parameters during tasks. However, no validation has been carried out for running gait analysis. The [...] Read more.
Running is the basis of many sports and has highly beneficial effects on health. To increase the understanding of running, DSPro® insoles were developed to collect running parameters during tasks. However, no validation has been carried out for running gait analysis. The aims of this study were to assess the test–retest reliability and criterion validity of running gait parameters from DSPro® insoles compared to a motion-capture system. Equipped with DSPro® insoles, a running gait analysis was performed on 30 healthy participants during overground and treadmill running using a motion-capture system. Using an intraclass correlation coefficient (ICC), the criterion validity and test–retest reliability of spatiotemporal parameters were calculated. The test–retest reliability shows moderate to excellent ICC values (ICC > 0.50) except for propulsion time during overground running at a fast speed with the motion-capture system. The criterion validity highlights a validation of running parameters regardless of speeds (ICC > 0.70). This present study validates the good criterion validity and test–retest reliability of DSPro® insoles for measuring spatiotemporal running gait parameters. Without the constraints of a 3D motion-capture system, such insoles seem to be helpful and relevant for improving the care management of active patients or following running performance in sports contexts. Full article
(This article belongs to the Special Issue Wearable Sensors for Human Health Monitoring and Analysis)
►▼ Show Figures

Figure 1

Figure 1
<p>DSPro<sup>®</sup> insole device.</p> Full article ">Figure 2
<p>Bland–Altman plot for all running parameters during fast (black points) and comfortable (red points) overground running. Solid line = mean, dashed line = ±1.96 SD.</p> Full article ">Figure 3
<p>Bland–Altman plot for all running parameters during fast (black points) and comfortable (red points) treadmill running. Solid line = mean, dashed line = ±1.96 SD.</p> Full article ">
10 pages, 892 KiB  
Article
Disturbances in Electrodermal Activity Recordings Due to Different Noises in the Environment
by Dindar S. Bari, Haval Y. Y. Aldosky, Christian Tronstad and Ørjan G. Martinsen
Sensors 2024, 24(16), 5434; https://doi.org/10.3390/s24165434 - 22 Aug 2024
Viewed by 452
Abstract
Electrodermal activity (EDA) is a widely used psychophysiological measurement in laboratory-based studies. In recent times, these measurements have seen a transfer from the laboratory to wearable devices due to the simplicity of EDA measurement as well as modern electronics. However, proper conditions for [...] Read more.
Electrodermal activity (EDA) is a widely used psychophysiological measurement in laboratory-based studies. In recent times, these measurements have seen a transfer from the laboratory to wearable devices due to the simplicity of EDA measurement as well as modern electronics. However, proper conditions for EDA measurement are recommended once wearable devices are used, and the ambient conditions may influence such measurements. It is not completely known how different types of ambient noise impact EDA measurement and how this translates to wearable EDA measurement. Therefore, this study explored the effects of various noise disturbances on the generation of EDA responses using a system for the simultaneous recording of all measures of EDA, i.e., skin conductance responses (SCRs), skin susceptance responses (SSRs), and skin potential responses (SPRs), at the same skin site. The SCRs, SSRs, and SPRs due to five types of noise stimuli at different sound pressure levels (70, 75, 80, 85, and 90 dB) were measured from 40 participants. The obtained results showed that EDA responses were generated at all levels and that the EDA response magnitudes were significantly (p < 0.001) influenced by the increasing noise levels. Different types of environmental noise may elicit EDA responses and influence wearable recordings outside the laboratory, where such noises are more likely than in standardized laboratory tests. Depending on the application, it is recommended to prevent these types of unwanted variation, presenting a challenge for the quality of wearable EDA measurement in real-world conditions. Future developments to shorten the quality gap between standardized laboratory-based and wearable EDA measurements may include adding microphone sensors and algorithms to detect, classify, and process the noise-related EDA. Full article
(This article belongs to the Special Issue Advanced Non-Invasive Sensors: Methods and Applications)
►▼ Show Figures

Figure 1

Figure 1
<p>Typical timeline for the EDA recordings.</p> Full article ">Figure 2
<p>Examples to illustrate how the onsets, peaks, and amplitudes of the SCRs, SSRs, and SPRs are specified.</p> Full article ">Figure 3
<p>(<b>a</b>) SCRs_Amp, * <span class="html-italic">p</span> &lt; 0.05 and ** <span class="html-italic">p</span> &lt; 0.005; (<b>b</b>) SPRs_Amp, * <span class="html-italic">p</span> &lt; 0.05 and ** <span class="html-italic">p</span> &lt; 0.005; and (<b>c</b>) SSRs_Amp, *** <span class="html-italic">p</span> &lt; 0.001 as a function of noise level; relation between noise levels and (<b>d</b>) mean values of SCRs_Amp, 95% CI from 0.8 to 1.5; (<b>e</b>) mean values of SPRs_Amp, 95% CI from −0.891 to −0.498; (<b>f</b>) SSRs_Amp, 95% CI from −0.320 to −0.063; (<b>g</b>) SCRs_Tris as a function of noise level, * <span class="html-italic">p</span> &lt; 0.01 and ** <span class="html-italic">p</span> &lt; 0.001; (<b>h</b>) SPRET as a function of noise level, * <span class="html-italic">p</span> &lt; 0.01 and ** <span class="html-italic">p</span> &lt; 0.001; relation between noise levels and (<b>i</b>) mean values of SCRs_Tris and noise level, 95% CI from 0.214 to 0.628; (<b>j</b>) mean values of SPRET and noise level, 95% CI from 4.914 to 9.677. In all boxplots, whiskers represent the maximum and minimum of the data.</p> Full article ">Figure A1
<p>A typical EDA waveform recorded from a participant shows variations in EDA responses following noise stimuli. SC = skin conductance, SS = skin susceptance, and SP = skin potential.</p> Full article ">
20 pages, 4393 KiB  
Article
Tool State Recognition Based on POGNN-GRU under Unbalanced Data
by Weiming Tong, Jiaqi Shen, Zhongwei Li, Xu Chu, Wenqi Jiang and Liguo Tan
Sensors 2024, 24(16), 5433; https://doi.org/10.3390/s24165433 - 22 Aug 2024
Viewed by 377
Abstract
Accurate recognition of tool state is important for maximizing tool life. However, the tool sensor data collected in real-life scenarios has unbalanced characteristics. Additionally, although graph neural networks (GNNs) show excellent performance in feature extraction in the spatial dimension of data, it is [...] Read more.
Accurate recognition of tool state is important for maximizing tool life. However, the tool sensor data collected in real-life scenarios has unbalanced characteristics. Additionally, although graph neural networks (GNNs) show excellent performance in feature extraction in the spatial dimension of data, it is difficult to extract features in the temporal dimension efficiently. Therefore, we propose a tool state recognition method based on the Pruned Optimized Graph Neural Network-Gated Recurrent Unit (POGNN-GRU) under unbalanced data. Firstly, design the Improved-Majority Weighted Minority Oversampling Technique (IMWMOTE) by introducing an adaptive noise removal strategy and improving the MWMOTE to alleviate the unbalanced problem of data. Subsequently, propose a POG graph data construction method based on a multi-scale multi-metric basis and a Gaussian kernel weight function to solve the problem of one-sided description of graph data under a single metric basis. Then, construct the POGNN-GRU model to deeply mine the spatial and temporal features of the data to better identify the state of the tool. Finally, validation and ablation experiments on the PHM 2010 and HMoTP datasets show that the proposed method outperforms the other models in terms of identification, and the highest accuracy improves by 1.62% and 1.86% compared with the corresponding optimal baseline model. Full article
(This article belongs to the Special Issue Fault Diagnosis and Prognosis for Electromechanical Actuators and Sensors)
►▼ Show Figures

Figure 1

Figure 1
<p>A POGNN-GRU-based model framework for tool state recognition under unbalanced data.</p> Full article ">Figure 2
<p>POGNN-GRU model framework.</p> Full article ">Figure 3
<p>(<b>a</b>) Wear variation curves of tools C1; (<b>b</b>) wear variation curves of tools C4; (<b>c</b>) wear variation curves of tools C6.</p> Full article ">Figure 4
<p>(<b>a</b>) Partial force data for single-cycle experiments during tool engagement; (<b>b</b>) Partial force data for single-cycle experiments during tool disengagement.</p> Full article ">Figure 5
<p>(<b>a</b>) Wear variation curves of T01; (<b>b</b>) wear variation curves of T02; (<b>c</b>) wear variation curves of T03.</p> Full article ">Figure 6
<p>Model training and testing results; (<b>a</b>) the result of PHM2010 dataset; (<b>b</b>) the result of HMoTP dataset.</p> Full article ">Figure 7
<p>Confusion Matrix, (<b>a</b>) Medium_Wear/Slight_Wear classification (no sampling) under PHM2010 dataset; (<b>b</b>) Medium_Wear/Severe_Wear classification (no sampling) under PHM2010 dataset; (<b>c</b>) Medium_Wear/Slight_Wear classification (with IMWMOTE) under PHM2010 dataset; (<b>d</b>) Medium_Wear/Severe_Wear classification (with IMWMOTE) under PHM2010 dataset; (<b>e</b>) Medium_Wear/Slight_Wear classification (no sampling) under HMoTP dataset; (<b>f</b>) Medium_Wear/Severe_Wear classification (no sampling) under HMoTP dataset; (<b>g</b>) Medium_Wear/Slight_Wear classification (with IMWMOTE) under HMoTP dataset; (<b>h</b>) Medium_Wear/Severe_Wear classification (with IMWMOTE) under HMoTP dataset.</p> Full article ">Figure 8
<p>The computational cost of different models.</p> Full article ">Figure 9
<p>The result of ablation experiments 1.</p> Full article ">
43 pages, 8972 KiB  
Review
Newest Methods and Approaches to Enhance the Performance of Optical Frequency-Domain Reflectometers
by Ivan A. Lobach, Andrei A. Fotiadi, Vasily A. Yatseev, Yuri A. Konstantinov, Fedor L. Barkov, D. Claude, Dmitry A. Kambur, Maxim E. Belokrylov, Artem T. Turov and Dmitry A. Korobko
Sensors 2024, 24(16), 5432; https://doi.org/10.3390/s24165432 - 22 Aug 2024
Viewed by 1237
Abstract
In this review, we summarize the latest advances in the design of optical frequency-domain reflectometers (OFDRs), digital signal processing, and sensors based on special optical fibers. We discuss state-of-the-art approaches to improving metrological characteristics, such as spatial resolution, SNR, dynamic range, and the [...] Read more.
In this review, we summarize the latest advances in the design of optical frequency-domain reflectometers (OFDRs), digital signal processing, and sensors based on special optical fibers. We discuss state-of-the-art approaches to improving metrological characteristics, such as spatial resolution, SNR, dynamic range, and the accuracy of determining back reflection coefficients. We also analyze the latest achievements in the OFDR-based sensors: the accuracy of spatial localization of the impact, the error in detecting temperatures, deformation, and other quantities, and the features of separate measurement of various physical quantities. We also pay attention to the trend of mutual integration of frequency-domain optical reflectometry methods with time-domain optical reflectometry, which provides completely new sensing possibilities. We believe that this review may be useful to engineers and scientists focused on developing a lab setup, complete measurement instrument, or sensing system with specific requirements. Full article
(This article belongs to the Special Issue Sensing Technologies and Optical Communication)
►▼ Show Figures

Figure 1

Figure 1
<p>The setup proposed in [<a href="#B1-sensors-24-05432" class="html-bibr">1</a>].</p> Full article ">Figure 2
<p>OFDR based on (<b>a</b>) Michelson and (<b>b</b>) Mach–Zehnder interferometer.</p> Full article ">Figure 3
<p>(<b>a</b>) OFDR setup based on Mach–Zehnder interferometer with two reference channels. (<b>b</b>) The output of this setup: without (orange) and with (blue) nonlinearity compensation. The orange plot represents the data obtained after FFT of the original signal. The blue plot shows the operation of one of the nonlinearity compensation algorithms, the data for which was obtained using an AUX. It is seen that the compensated data differ in that the beginning and end of the line are clear peaks; while in the raw data, rather wide areas can be observed instead of fiber ends, where information about the exact boundaries of the FUT is lost.</p> Full article ">Figure 4
<p>The laser stabilization proposed in [<a href="#B36-sensors-24-05432" class="html-bibr">36</a>].</p> Full article ">Figure 5
<p>Standard algorithm for obtaining temperatures and strains in OFDR.</p> Full article ">Figure 6
<p>Evolution of the temperature field obtained via OFDR means without the algorithm proposed by Sweeney and Petry (<b>left</b>) and with it (<b>right</b>).</p> Full article ">Figure 7
<p>The algorithm proposed in [<a href="#B61-sensors-24-05432" class="html-bibr">61</a>].</p> Full article ">Figure 8
<p>The setup proposed in [<a href="#B67-sensors-24-05432" class="html-bibr">67</a>].</p> Full article ">Figure 9
<p>Estimation of the possibility of temperatures and strains extracting from the SNR of the real part of the OFDR spectrum.</p> Full article ">Figure 10
<p>OFDR schematics presented in [<a href="#B73-sensors-24-05432" class="html-bibr">73</a>].</p> Full article ">Figure 11
<p>Multilayer perceptron used in [<a href="#B73-sensors-24-05432" class="html-bibr">73</a>] to improve OFDR sensing performance.</p> Full article ">Figure 12
<p>OFDR using vector network analyzer (VNA) proposed in [<a href="#B78-sensors-24-05432" class="html-bibr">78</a>].</p> Full article ">Figure 13
<p>Results obtained in [<a href="#B82-sensors-24-05432" class="html-bibr">82</a>] for OFDR when detecting a vibration.</p> Full article ">Figure 14
<p>ACF presented in [<a href="#B83-sensors-24-05432" class="html-bibr">83</a>].</p> Full article ">Figure 15
<p>Schematic diagram of the setup proposed in [<a href="#B88-sensors-24-05432" class="html-bibr">88</a>]. WDM: wavelength division multiplexer.</p> Full article ">Figure 16
<p>Two-dimensional representation of predicted and target values and linear regression for temperature (<b>left</b>) and strain (<b>right</b>) data.</p> Full article ">Figure 17
<p>Schematic diagram of the TGD-OFDR system. FG: function generator; AOM: acousto-optic modulator; BPD: balanced photodetector; ADC: analog-to-digital converter; PC: personal computer [<a href="#B90-sensors-24-05432" class="html-bibr">90</a>].</p> Full article ">Figure 18
<p>Frequencies of the beat signals received via a balanced photodetector (BPD) (<b>upper picture</b>), and the two equivalent references (<b>lower picture</b>) [<a href="#B90-sensors-24-05432" class="html-bibr">90</a>]. Pink line designates the digitally synthesized shifted probe signal.</p> Full article ">Figure 19
<p>Schematic of the pulse generation and frequency sweep approach [<a href="#B94-sensors-24-05432" class="html-bibr">94</a>].</p> Full article ">Figure 20
<p>RBS for initial conditions at a specific fiber location for the case of with (red dashed line) and without (blue line) FUT deformation. Adapted from [<a href="#B95-sensors-24-05432" class="html-bibr">95</a>].</p> Full article ">Figure 21
<p>OFDR with a combined main and AUX interferometer. Adapted from [<a href="#B106-sensors-24-05432" class="html-bibr">106</a>].</p> Full article ">Figure 22
<p>OFDR with reflector in the main interferometer. Adapted from [<a href="#B108-sensors-24-05432" class="html-bibr">108</a>].</p> Full article ">Figure 23
<p>The operating principle of OFDR based on a self-scanning fiber laser. (<b>a</b>) Laser radiation, consisting of a sequence of pulses with a fixed discrete frequency entering the main interferometer; (<b>b</b>) output of the interferometer, the sequence of pulses acquires an amplitude envelope; and (<b>c</b>) the extracted envelope has equidistant frequency samples.</p> Full article ">
19 pages, 7679 KiB  
Article
An Introduction to Ventra: A Programmable Abdominal Phantom for Training, Educational, Research, and Development Purposes
by Salar Tayebi, Robert Wise, Ashkan Zarghami, Wojciech Dabrowski, Manu L. N. G. Malbrain and Johan Stiens
Sensors 2024, 24(16), 5431; https://doi.org/10.3390/s24165431 - 22 Aug 2024
Viewed by 630
Abstract
Background: Intra-abdominal pressure (IAP) is a critical parameter in the care of critically ill patients, as elevated IAP can lead to reduced cardiac output and organ perfusion, potentially resulting in multiple organ dysfunction and failure. The current gold standard for measuring IAP is [...] Read more.
Background: Intra-abdominal pressure (IAP) is a critical parameter in the care of critically ill patients, as elevated IAP can lead to reduced cardiac output and organ perfusion, potentially resulting in multiple organ dysfunction and failure. The current gold standard for measuring IAP is an indirect technique via the bladder. According to the Abdominal Compartment Society’s Guidelines, new measurement methods/devices for IAP must be validated against the gold standard. Objectives: This study introduces Ventra, an abdominal phantom designed to simulate different IAP levels, abdominal compliance, respiration-related IAP variations, and bladder dynamics. Ventra aims to facilitate the development and validation of new IAP measurement devices while reducing reliance on animal and cadaveric studies. Additionally, it offers potential applications in training and education for biomedical engineering students. This study provides a thorough explanation on the phantom’s design and fabrication, which provides a low-cost solution for advancing IAP measurement research and education. The design concept, technical aspects, and a series of validation experiments determining whether Ventra is a suitable tool for future research are presented in this study. Methods: Ventra’s performance was evaluated through a series of validation tests using a pressure gauge and two intra-gastric (Spiegelberg and CiMON) and two intra-bladder (Accuryn and TraumaGuard) pressure measurement devices. The mean and standard deviation of IAP recordings by each device were investigated. Bland–Altman analysis was used to evaluate bias, precision, limits of agreement, and percentage error for each system. Concordance analysis was performed to assess the ability of Ventra in tracking IAP changes. Results: The phantom demonstrated excellent agreement with reference pressure measurements, showing an average bias of 0.11 ± 0.49 mmHg. A concordance coefficient of 100% was observed for the phantom as well. Ventra accurately simulated different abdominal compliances, with higher IAP values resulting in lower compliance. Abdominal volume changes showed a bias of 0.08 ± 0.07 L/min, and bladder fill volume measurements showed an average difference of 0.90 ± 4.33 mL for volumes ranging from 50 to 500 mL. Conclusion: The validation results were in agreement with the research guidelines of the world abdominal society. Ventra is a reliable tool that will facilitate the development and validation of new IAP measurement devices. It is an effective educational tool for biomedical engineering students as well. Full article
(This article belongs to the Section Biomedical Sensors)
►▼ Show Figures

Graphical abstract

Graphical abstract
Full article ">Figure 1
<p>An illustration of the abdominal compartment and abdominal wall of the phantom. (<b>a</b>) Two sub-compartments were placed inside the modeled abdominal compartment to represent the bladder and stomach. Two intra-bladder and two intra-gastric pressure measurement instruments were connected to the abdominal compartment as well. (<b>b</b>) An artificial abdominal wall including skin, fat, muscle, and peritoneum was placed on top of the abdominal compartment. By connecting the abdominal wall to the ground surface by extension springs, the abdominal compliance can be adjusted further.</p> Full article ">Figure 2
<p>The computer-aided design (CAD) of the enclosure of the phantom.</p> Full article ">Figure 3
<p>The IAP control module of the GUI. The IAP can be increased, decreased, and kept constant through the IAP control table. An automatic IAP simulation is also designed to deliver the requested IAP value.</p> Full article ">Figure 4
<p>The phantom setup. The phantom, its control unit, and its GUI in addition to the intra-bladder and intra-gastric measurement devices measuring an IAP of 7 mmHg.</p> Full article ">Figure 5
<p>The mean and standard deviation of the IAP measurements at 0, 5, 10, 15, 20, and 25 mmHg. The error bars are the 95% confidence interval, the bottom and top of the box are the 25th and 75th percentiles, respectively, the line inside the box is the 50th percentile (median), and any outliers are shown in red.</p> Full article ">Figure 6
<p>Bland–Altman results of the simulated IAP. Ventra showed a slight bias of +0.11 mmHg when compared with the pressure gauge.</p> Full article ">Figure 7
<p>The concordance analysis result. A concordance coefficient of 100% was observed for Ventra, which shows the capability of this abdominal phantom in tracking IAP changes (the blue region indicated the excluded data).</p> Full article ">Figure 8
<p>Compliance maneuver of the phantom at high-, medium-, and low-compliance conditions. A relatively high compliance can be observed at IAP values lower than 10 mmHg. However, at IAPs higher than 10 mmHg, the impact of the springs becomes more dominant, which results in reduced compliance.</p> Full article ">Figure 9
<p>Mean and standard deviation of ΔIAV recorded by the gas flow sensor. The measurements at four different flow rates are compared with the reference measurements by an analog flow meter.</p> Full article ">Figure 10
<p>Bland–Altman results of ΔIAV. A relatively small bias was noticed between Ventra and the analog flow meter.</p> Full article ">Figure 11
<p>Bladder fluid volume measurement results. The results show an accurate bladder fluid volume measurement.</p> Full article ">Figure 12
<p>Bland–Altman analysis of the bladder fill volume. As already noticed, an underestimation happens for volumes smaller than 200 mL. However, at higher volumes the underestimation becomes a slight overestimation.</p> Full article ">Figure 13
<p>The IAP measurements by the (<b>a</b>) Accuryn and (<b>b</b>) TraumaGuard catheters in Ventra at different bladder fill volumes.</p> Full article ">Figure 14
<p>Dynamic IAP changes due to respiration with I:E set at 1:1. IAP changes due to respiration with an abdominal pressure variation of (<b>a</b>) 10%, (<b>b</b>) 20%, and (<b>c</b>) 30%. The end-inspiration (IAP<sub>ei</sub>), end-expiration IAP (IAP<sub>ee</sub>), and ΔIAP are illustrated as well.</p> Full article ">
15 pages, 5898 KiB  
Article
Metamaterial Broadband Absorber Induced by Synergistic Regulation of Temperature and Electric Field and Its Optical Switching Application
by Rundong Yang, Yun Liu and Xiangfu Wang
Sensors 2024, 24(16), 5430; https://doi.org/10.3390/s24165430 - 22 Aug 2024
Viewed by 533
Abstract
Nowadays, metamaterial absorbers still suffer from limited bandwidth, poor bandwidth scalability, and insufficient modulation depth. In order to solve this series of problems, we propose a metamaterial absorber based on graphene, VO2, gallium silver sulfide, and gold-silver alloy composites with dual-control [...] Read more.
Nowadays, metamaterial absorbers still suffer from limited bandwidth, poor bandwidth scalability, and insufficient modulation depth. In order to solve this series of problems, we propose a metamaterial absorber based on graphene, VO2, gallium silver sulfide, and gold-silver alloy composites with dual-control modulation of temperature and electric field. Then we further investigate the optical switching performance of this absorber in this work. Our proposed metamaterial absorber has the advantages of broad absorption bandwidth, sufficient modulation depth, and good bandwidth scalability all together. Unlike the single inspired layer of previous designs, we innovatively adopted a multi-layer excitation structure, which can realize the purpose of absorption and bandwidth width regulation by a variety of means. Combined with the finite element analysis method, our proposed metamaterial absorber has excellent bandwidth scalability, which can be tuned from 2.7 THz bandwidth to 12.1 THz bandwidth by external electrothermal excitation. Meanwhile, the metamaterial absorber can also dynamically modulate the absorption from 3.8% to 99.8% at a wide incidence angle over the entire range of polarization angles, suggesting important potential applications in the field of optical switching in the terahertz range. Full article
(This article belongs to the Section Sensor Materials)
►▼ Show Figures

Figure 1

Figure 1
<p>(<b>a</b>) Cross-section of the MA unit structure. (<b>b</b>) Structural decomposition of the MA. (<b>c</b>) Planar structure of the graphene layer of the MA. (<b>d</b>) Planar structure of the VO<sub>2</sub>-gold-silver alloy layer of the MA. (<b>e</b>) Planar structure of the MA VO<sub>2</sub>-gallium sulfide silver planar structure of the MA.</p> Full article ">Figure 2
<p>Multiple reflection interference equivalent model of MA.</p> Full article ">Figure 3
<p>Relationship between simulated and theoretically calculated absorption spectra of the MA at (<b>a</b>) <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 200 S/m, and (<b>b</b>) <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 20,000 S/m.</p> Full article ">Figure 4
<p>Absorption spectra of MA with and without AgGaS<sub>2</sub> at (<b>a</b>) <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 200 S/m, and (<b>b</b>) <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 20,000 S/m.</p> Full article ">Figure 5
<p>When MA graphene surface <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 200 S/m, and 4.5 THz wave incidence (<b>a</b>) Graphene layer electric field strength <span class="html-italic">|E|</span> distribution, (<b>b</b>) VO<sub>2</sub> layer electric field strength <span class="html-italic">|E|</span> distribution, (<b>c</b>) Gallium silver sulfur layer electric field strength <span class="html-italic">|E|</span> distribution, (<b>d</b>) Gold substrate surface electric field strength <span class="html-italic">|E|</span> distribution, and (<b>e</b>) <span class="html-italic">y</span> = 0 at the magnetic field strength <span class="html-italic">|H|</span> distribution in the <span class="html-italic">xoz</span> plane.</p> Full article ">Figure 6
<p>When MA graphene surface <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 20,000 S/m, MA graphene layer surface electric field strength <span class="html-italic">|E|</span> distribution in the case of (<b>a</b>) 5 THz, (<b>b</b>) 10 THz, and (<b>c</b>) 15 THz incident wave.</p> Full article ">Figure 7
<p>Absorption spectrum of the MA at <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 200 S/m as a function of (<b>a</b>) <span class="html-italic">d</span><sub>1</sub>, (<b>b</b>) <span class="html-italic">R</span><sub>1</sub>, and (<b>c</b>) <span class="html-italic">R</span><sub>3</sub>.</p> Full article ">Figure 8
<p>Absorption spectra of the MA at <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 20,000 S/m as a function of (<b>a</b>) <span class="html-italic">R</span><sub>3</sub> and (<b>b</b>) <span class="html-italic">R</span><sub>4</sub>.</p> Full article ">Figure 9
<p>Absorption spectra of the MA as a function of (<b>a</b>) different Fermi energy levels when <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 200 S/m, (<b>b</b>) different Fermi energy levels when <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 20,000 S/m, (<b>c</b>) different relaxation times when <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 200 S/m, (<b>d</b>) different relaxation times when <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 20,000 S/m, and (<b>e</b>) different temperatures when <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">τ</span> = 0.058 ps.</p> Full article ">Figure 10
<p>Absorption spectrum of the MA at different polarization angles for (<b>a</b>) <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 20,000 S/m when <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, and (<b>b</b>) <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 200 S/m when <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">E<sub>f</sub></span> = 1.25 eV.</p> Full article ">Figure 11
<p>Absorption spectra of (<b>a</b>) TE wave at different incidence cases with MA at <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 200 S/m, (<b>b</b>) TE wave at different incidence cases with MA at <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 20,000 S/m, (<b>c</b>) TM wave at different incidence cases with MA at <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 200 S/m, and (<b>d</b>) TE wave at different incidence cases with MA at <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">E<sub>f</sub></span> = 1.25 eV, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 20,000 S/m.</p> Full article ">Figure 12
<p>(<b>a</b>) Absorption spectra of Fermi energy levels <span class="html-italic">E<sub>f</sub></span> = 0 eV and <span class="html-italic">E<sub>f</sub></span> = 1.25 eV when the MA <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 200 S/m. (<b>b</b>) Absorption at 3.2 THz, 4.2 THz, and 5.2 THz for different <span class="html-italic">E<sub>f</sub></span> when the MA <span class="html-italic">τ</span> = 0.058 ps, <span class="html-italic">σ<sub>vo</sub></span><sub>2</sub> = 200 S/m.</p> Full article ">
18 pages, 5410 KiB  
Article
Transformer and Adaptive Threshold Sliding Window for Improving Violence Detection in Videos
by Fernando J. Rendón-Segador, Juan A. Álvarez-García and Luis M. Soria-Morillo
Sensors 2024, 24(16), 5429; https://doi.org/10.3390/s24165429 - 22 Aug 2024
Viewed by 509
Abstract
This paper presents a comprehensive approach to detect violent events in videos by combining CrimeNet, a Vision Transformer (ViT) model with structured neural learning and adversarial regularization, with an adaptive threshold sliding window model based on the Transformer architecture. CrimeNet demonstrates exceptional performance [...] Read more.
This paper presents a comprehensive approach to detect violent events in videos by combining CrimeNet, a Vision Transformer (ViT) model with structured neural learning and adversarial regularization, with an adaptive threshold sliding window model based on the Transformer architecture. CrimeNet demonstrates exceptional performance on all datasets (XD-Violence, UCF-Crime, NTU-CCTV Fights, UBI-Fights, Real Life Violence Situations, MediEval, RWF-2000, Hockey Fights, Violent Flows, Surveillance Camera Fights, and Movies Fight), achieving high AUC ROC and AUC PR values (up to 99% and 100%, respectively). However, the generalization of CrimeNet to cross-dataset experiments posed some problems, resulting in a 20–30% decrease in performance, for instance, training in UCF-Crime and testing in XD-Violence resulted in 70.20% in AUC ROC. The sliding window model with adaptive thresholding effectively solves these problems by automatically adjusting the violence detection threshold, resulting in a substantial improvement in detection accuracy. By applying the sliding window model as post-processing to CrimeNet results, we were able to improve detection accuracy by 10% to 15% in cross-dataset experiments. Future lines of research include improving generalization, addressing data imbalance, exploring multimodal representations, testing in real-world applications, and extending the approach to complex human interactions. Full article
(This article belongs to the Collection Advances in Deep-Learning-Based Sensing, Imaging, and Video Processing)
►▼ Show Figures

Figure 1

Figure 1
<p>Adaptive threshold sliding window model architecture. It uses an Embedding layer together with a MultiHead Attention layer following the philosophy of the Transformer models. This architecture takes advantage of the strengths of the Transformer models to establish relationships between the different elements of a sequence of data in a one-dimensional vector.</p> Full article ">Figure 2
<p>Synthetic data labeling process using autoencoder plus K-Means. The autoencoder takes the synthetically generated binary sequences as input and encodes them. From these encodings, the K-Means algorithm groups the corresponding sequences into two clusters and labels them according to the cluster to which they belong after the application of K-Means.</p> Full article ">Figure 3
<p>Complete system that integrates CrimeNet with the adaptive threshold sliding window model to correct false positives and negatives that may be generated by CrimeNet.</p> Full article ">Figure 4
<p>CrimeNet cross-dataset comparison with initial datasets (NTU CCTV-Fights, UBI-Fights, XD-Violence, and UCF-Crime).</p> Full article ">Figure 5
<p>False positive and negative percentages in cross-datasets for NTU CCTV Fights.</p> Full article ">Figure 6
<p>False positive and negative percentages in cross-datasets for UBI Fights.</p> Full article ">Figure 7
<p>AUC ROC and AUC PR curve for the adaptive threshold sliding window model.</p> Full article ">Figure 8
<p>Attention maps of the adaptive threshold sliding window model for three binary sequences of length ten. The first attention map corresponds to a sequence whose values are all zeros, the second one corresponds to the sequence: [0, 0, 0, 0, 1, 1, 1, 1, 0, 1, 0] and the third one to a sequence whose values are all ones.</p> Full article ">Figure 9
<p>AUC ROC and AUC PR for different percentages of violence in cross-dataset of NTU CCTV Fights-UBI Fights.</p> Full article ">Figure 10
<p>AUC ROC and AUC PR for different percentages of violence in cross-dataset of UBI Fights-NTU CCTV Fights.</p> Full article ">Figure 11
<p>Elimination of false positives in a fragment of the movie <span class="html-italic">John Wick 2</span>. <b>Top row</b>: CrimeNet model predictions with one false positive. <b>Bottom row</b>: post-processing with the adaptive threshold sliding window model.</p> Full article ">Figure 12
<p>K-Means vs. adaptive threshold sliding window in <span class="html-italic">John Wick 2</span> cross-dataset comparison.</p> Full article ">
14 pages, 796 KiB  
Article
Independent Vector Analysis for Feature Extraction in Motor Imagery Classification
by Caroline Pires Alavez Moraes, Lucas Heck dos Santos, Denis Gustavo Fantinato, Aline Neves and Tülay Adali
Sensors 2024, 24(16), 5428; https://doi.org/10.3390/s24165428 - 22 Aug 2024
Viewed by 488
Abstract
Independent vector analysis (IVA) can be viewed as an extension of independent component analysis (ICA) to multiple datasets. It exploits the statistical dependency between different datasets through mutual information. In the context of motor imagery classification based on electroencephalogram (EEG) signals for the [...] Read more.
Independent vector analysis (IVA) can be viewed as an extension of independent component analysis (ICA) to multiple datasets. It exploits the statistical dependency between different datasets through mutual information. In the context of motor imagery classification based on electroencephalogram (EEG) signals for the brain–computer interface (BCI), several methods have been proposed to extract features efficiently, mainly based on common spatial patterns, filter banks, and deep learning. However, most methods use only one dataset at a time, which may not be sufficient for dealing with a multi-source retrieving problem in certain scenarios. From this perspective, this paper proposes an original approach for feature extraction through multiple datasets based on IVA to improve the classification of EEG-based motor imagery movements. The IVA components were used as features to classify imagined movements using consolidated classifiers (support vector machines and K-nearest neighbors) and deep classifiers (EEGNet and EEGInception). The results show an interesting performance concerning the clustering of MI-based BCI patients, and the proposed method reached an average accuracy of 86.7%. Full article
(This article belongs to the Section Biomedical Sensors)
►▼ Show Figures

Figure 1

Figure 1
<p>Procedure description to obtain the IVA matrices <math display="inline"><semantics> <msubsup> <mi mathvariant="bold">W</mi> <mrow> <mi>c</mi> </mrow> <mrow> <mo>[</mo> <mi>k</mi> <mo>]</mo> </mrow> </msubsup> </semantics></math> for each class based on the training data.</p> Full article ">Figure 2
<p>Procedure description for the <span class="html-italic">k</span>-th subject used in training and test datasets.</p> Full article ">Figure 3
<p>Performance analysis of IVAS and IVAK concerning IVA initialization based on the KDE for subjects from Dataset4a. (<b>a</b>) Dataset4a with IVAS; (<b>b</b>) Dataset4a with IVAK.</p> Full article ">Figure 4
<p>IVAS and IVAK performance analysis with respect to the number of EEG channels.</p> Full article ">Figure 5
<p>Examples of SCV covariance matrices obtained through IVA for the DS4a considering right hand and right foot movements.</p> Full article ">
16 pages, 844 KiB  
Article
Cascade Proportional–Integral Control Design and Affordable Instrumentation System for Enhanced Performance of Electrolytic Dry Cells
by Saulo N. Matos, Gemírson de Paula dos Reis, Elisângela M. Leal, Robson L. Figueiredo, Thiago A. M. Euzébio and Alan K. Rêgo Segundo
Sensors 2024, 24(16), 5427; https://doi.org/10.3390/s24165427 - 22 Aug 2024
Viewed by 535
Abstract
In this paper, we present a cost-effective system for monitoring and controlling alkaline electrolyzers, intending to improve hydrogen gas production on a laboratory scale. Our work includes two main innovations. Firstly, we suggest an approach to calibrate a standard air flow meter to [...] Read more.
In this paper, we present a cost-effective system for monitoring and controlling alkaline electrolyzers, intending to improve hydrogen gas production on a laboratory scale. Our work includes two main innovations. Firstly, we suggest an approach to calibrate a standard air flow meter to accurately measure the flow of hydrogen-rich gas from electrolyzers, improving measurement accuracy while keeping costs low. Secondly, we introduce a unique cascade control method to manage hydrogen-rich gas production in the electrolyzer, ensuring precise control over gas flow rates. By combining affordable, energy-efficient devices with a PI control system, we achieve efficient gas production through electrolysis, replacing manual control approaches. Experimental results confirm the effectiveness of our cascade control method, demonstrating stable operation with minimal errors. These results provide a foundation for further research into control strategies to enhance the performance of electrolytic cells. Full article
(This article belongs to the Collection Sensors and Intelligent Control Systems)
►▼ Show Figures

Figure 1

Figure 1
<p>Experiment setup for electrolytic cell.</p> Full article ">Figure 2
<p>Schematic diagram of instrumentation and control system.</p> Full article ">Figure 3
<p>Measurement chain.</p> Full article ">Figure 4
<p>Block diagram for a cascade control approach. Adapted from [<a href="#B38-sensors-24-05427" class="html-bibr">38</a>].</p> Full article ">Figure 5
<p>Open-loop step response for the current control system. The PWM duty cycle step is shown in red, and the current response is in blue.</p> Full article ">Figure 6
<p>Open-loop step response for the gas flow and current control system. The current is shown in red, and the mass flow rate response is shown in blue.</p> Full article ">Figure 7
<p>System calibration using a linear regression model. The units on the vertical axis represent ADC counts.</p> Full article ">Figure 8
<p>Measured variables by the instrumentation system. The superior plots from left to right show the current and voltage, respectively. The inferior plots show the mass flow and the electrolyzer temperature.</p> Full article ">Figure 9
<p>Process Reaction Curve method. The mass flow rate is indicated by the blue signal, the green line marks the step instant, and the red line denotes the tangent at the inflection point.</p> Full article ">Figure 10
<p>Nyquist plot for the open inner loop with maximum sensitivity <math display="inline"><semantics> <msub> <mi>M</mi> <mrow> <mi>s</mi> <mn>1</mn> </mrow> </msub> </semantics></math> equals 1.28.</p> Full article ">Figure 11
<p>Nyquist plot for the open outer loop with maximum sensitivity <math display="inline"><semantics> <msub> <mi>M</mi> <mrow> <mi>s</mi> <mn>2</mn> </mrow> </msub> </semantics></math> equals 1.17.</p> Full article ">Figure 12
<p>Gas flow control behavior. The set-point is shown in red, while the measured values are in blue.</p> Full article ">Figure 13
<p>Current control behavior. The set-point is shown in red, while the measured values are in blue.</p> Full article ">Figure 14
<p>The error curve during the experiment with set-point variation.</p> Full article ">
18 pages, 11471 KiB  
Article
Advanced Sensing System for Sleep Bruxism across Multiple Postures via EMG and Machine Learning
by Jahan Zeb Gul, Noor Fatima, Zia Mohy Ud Din, Maryam Khan, Woo Young Kim and Muhammad Muqeet Rehman
Sensors 2024, 24(16), 5426; https://doi.org/10.3390/s24165426 - 22 Aug 2024
Viewed by 717
Abstract
Diagnosis of bruxism is challenging because not all contractions of the masticatory muscles can be classified as bruxism. Conventional methods for sleep bruxism detection vary in effectiveness. Some provide objective data through EMG, ECG, or EEG; others, such as dental implants, are less [...] Read more.
Diagnosis of bruxism is challenging because not all contractions of the masticatory muscles can be classified as bruxism. Conventional methods for sleep bruxism detection vary in effectiveness. Some provide objective data through EMG, ECG, or EEG; others, such as dental implants, are less accessible for daily practice. These methods have targeted the masseter as the key muscle for bruxism detection. However, it is important to consider that the temporalis muscle is also active during bruxism among masticatory muscles. Moreover, studies have predominantly examined sleep bruxism in the supine position, but other anatomical positions are also associated with sleep. In this research, we have collected EMG data to detect the maximum voluntary contraction of the temporalis and masseter muscles in three primary anatomical positions associated with sleep, i.e., supine and left and right lateral recumbent positions. A total of 10 time domain features were extracted, and six machine learning classifiers were compared, with random forest outperforming others. The models achieved better accuracies in the detection of sleep bruxism with the temporalis muscle. An accuracy of 93.33% was specifically found for the left lateral recumbent position among the specified anatomical positions. These results indicate a promising direction of machine learning in clinical applications, facilitating enhanced diagnosis and management of sleep bruxism. Full article
(This article belongs to the Section Biomedical Sensors)
►▼ Show Figures

Figure 1

Figure 1
<p>The overall methodology of this research. (<b>A</b>) shows a person doing jaw clenching during sleep. (<b>B</b>) shows the data acquisition using the EMG module from MP36 Biopac. The surface electrodes are placed on the two concerned muscles, i.e., the temporalis and masseter muscles. (<b>C</b>) The data were acquired from the patient lying in the supine and left and right lateral recumbent positions, following this paradigm with initial, intermediate, and final rest periods of 5 s each, and motor activities of 10 s. The motor activities performed in this experiment comprised all the possible motor movements associated with the two concerned muscles.</p> Full article ">Figure 2
<p>The EMG surface electrodes were placed on the temporalis and masseter muscles with common ground electrodes placed on the frontal bone. The temporalis muscle was set as channel 1 while the masseter muscle was set as channel 2. The signal was acquired using BIOPAC MP36, which mainly comprises an instrumentation amplifier to capture the signal with a high pass filter of 30 Hz and a low pass filter of 1000 Hz. The data acquired were displayed on the PC in real-time from two channels, which were further stored to perform signal processing and feature extraction.</p> Full article ">Figure 3
<p>The subjects lay first in the supine position and then in the left and right lateral recumbent positions and performed bruxism activities associated with teeth grinding, jaw clenching, and teeth tapping as indicated with arrows while EMG electrodes were placed on the temporalis and masseter muscles on the right side of the patient.</p> Full article ">Figure 4
<p>The MATLAB plot of EMG acquired in 65 s from the temporalis muscle and masseter muscles demonstrates the segmented blocks based on the sampling rate of 2000 Hz. Each rest block i.e., initial, intermediate, and final, is 5 s in length, and each motor activity block is 10 s in length.</p> Full article ">Figure 5
<p>The signal in MATLAB was subjected to pre-processing before feature selection and feature extraction. The features were selected based on extensive literature research related to our application. A total of 10 features were selected, which were then fed as input to the machine learning classifiers. The Adaptive Synthetic Sampling Approach was applied to handle the imbalance of binary classes. The features were then reduced using principal component analysis and stratified 10-fold cross-validation was applied to divide the dataset into split, test, and valid sets. The models were then trained, and their performance was evaluated for bruxism detection.</p> Full article ">Figure 6
<p>The CumSum plots for bilateral comparison of the temporalis muscle in the left lateral recumbent position.</p> Full article ">Figure 7
<p>The CumSum plots for bilateral comparison of the temporalis muscle in the right lateral recumbent position.</p> Full article ">Figure 8
<p>The CumSum plots for the bilateral comparison of the masseter muscle in the left lateral recumbent position.</p> Full article ">Figure 9
<p>The CumSum plots for bilateral comparison of the masseter muscle in the right lateral recumbent position.</p> Full article ">Figure 10
<p>The power spectrum plot for the emporalis muscle in the supine position, and masseter muscle in the supine position.</p> Full article ">Figure 11
<p>The power spectrum plot for the temporalis muscle in the left lateral recumbent position, and the masseter muscle in the left lateral recumbent position.</p> Full article ">Figure 12
<p>The power spectrum plot for the temporalis muscle in the right lateral recumbent position, and the masseter muscle in the left lateral recumbent position.</p> Full article ">Figure 13
<p>Comparison of accuracies of machine learning models at supine left lateral recumbent and right lateral recumbent positions for the masticatory muscles: (<b>a</b>) temporalis; (<b>b</b>) masseter.</p> Full article ">Figure 14
<p>The confusion matrices from the support vector machine model for (<b>a</b>) the temporalis muscle in a supine position, (<b>b</b>) the temporalis muscle in a left lateral recumbent position, and (<b>c</b>) the temporalis muscle in a right lateral recumbent position.</p> Full article ">
13 pages, 10016 KiB  
Article
ZIF-8-Derived Multifunctional Triethylamine Sensor
by Shuo Xiao, Zheng Jiao and Xuechun Yang
Sensors 2024, 24(16), 5425; https://doi.org/10.3390/s24165425 - 22 Aug 2024
Viewed by 416
Abstract
Triethylamine (TEA) is a typical volatile organic compound (VOC) widely present in air and water, produced in industrial production activities, with high toxicity and great harm. Fluorescence detection and resistive sensing are effective methods for detecting pollutants. Here, In-doped interpenetrating twin ZIF-8 and [...] Read more.
Triethylamine (TEA) is a typical volatile organic compound (VOC) widely present in air and water, produced in industrial production activities, with high toxicity and great harm. Fluorescence detection and resistive sensing are effective methods for detecting pollutants. Here, In-doped interpenetrating twin ZIF-8 and its annealed derivatives have been successfully designed and prepared as a multifunctional TEA sensor. On the one hand, ZIF-8-In exhibits excellent fluorescence emission enhancement at 450 nm in a dose-dependent manner to TEA in water within the concentration range of 1–100 ppm, with a detection limit as low as 1 ppm. On the other hand, the annealed ZIF-8-In derivative is ZnO/In2O3 with a porous hierarchical structure, which is a perfect sensitive material for manufacturing gas sensors. Within the concentration range of 1–100 ppm, the ZnO/In2O3 gas sensor has a high response for 100 ppm TEA, reaching 107.7 (Ra/Rg), and can detect TEA gas as low as 1 ppm. Furthermore, the response of ZnO/In2O3 sensors to TEA is at least 10 times that of the other four VOC gases, demonstrating excellent gas selectivity. This multifunctional sensor can adapt to complex detection situations, demonstrating good application prospects. Full article
(This article belongs to the Section Chemical Sensors)
►▼ Show Figures

Figure 1

Figure 1
<p>(<b>a</b>,<b>b</b>) The SEM images of ZIF-8, (<b>c</b>) the SEM image of ZIF-8-In, (<b>d</b>) the BET results of ZIF-8-In, (<b>e</b>) the XRD patterns of ZIF-8 and ZIF-8-In.</p> Full article ">Figure 2
<p>(<b>a</b>) The survey spectra of ZIF-8 and ZIF-8-In, the high-resolution (<b>b</b>) C1s, (<b>c</b>) Zn2p, and (<b>d</b>) In3d spectra of ZIF-8-In.</p> Full article ">Figure 3
<p>(<b>a</b>) The emission spectra of ZIF-8 and ZIF-8-In excited by 365 nm, (<b>b</b>) the emission spectra, and (<b>c</b>) normalized emission spectra of ZIF-8-In under different excitation wavelengths, (<b>d</b>) the trend of emission peak position of ZIF-8-In with the change in excitation wavelength.</p> Full article ">Figure 4
<p>(<b>a</b>) The change in fluorescence intensity of ZIF-8-In with the addition of 1–100 ppm TEA, (<b>b</b>) relationships between the concentration of TEA and emission intensities of ZIF-8-In, the selectivity of ZIF-8-In for (<b>c</b>) 100 ppm single pollutants and for (<b>d</b>) 100 ppm mixed pollutants.</p> Full article ">Figure 5
<p>(<b>a</b>) The SEM image of ZnO/In<sub>2</sub>O<sub>3</sub>, (<b>b</b>–<b>e</b>) the TEM images of ZnO/In<sub>2</sub>O<sub>3</sub>, (<b>f</b>) the XRD result of ZnO/In<sub>2</sub>O<sub>3</sub>.</p> Full article ">Figure 6
<p>The survey XPS spectra of (<b>a</b>) ZIF-8-In and (<b>b</b>) ZnO/In<sub>2</sub>O<sub>3</sub>, the high-resolution O1s spectra of (<b>c</b>) ZIF-8-In and (<b>d</b>) ZnO/In<sub>2</sub>O<sub>3</sub>.</p> Full article ">Figure 7
<p>Manufacturing diagram of ZnO/In<sub>2</sub>O<sub>3</sub> gas sensor, (<b>a</b>) structure diagram of gas sensor and (<b>b</b>) the photo of gas sensor measurement system.</p> Full article ">Figure 8
<p>The gas response of the ZnO and ZnO/In<sub>2</sub>O<sub>3</sub> sensors towards 100 ppm TEA at different operating temperatures.</p> Full article ">Figure 9
<p>(<b>a</b>) The response of the ZnO/In<sub>2</sub>O<sub>3</sub> sensor to 1–100 ppm TEA at 200 °C, (<b>b</b>) the relationships between TEA concentration and response of ZnO/In<sub>2</sub>O<sub>3</sub>, (<b>c</b>) the response repeatability curve of ZnO/In<sub>2</sub>O<sub>3</sub> sensor to 100 ppm TEA at 200 °C, (<b>d</b>) the response of ZnO/In<sub>2</sub>O<sub>3</sub> sensor to 100 ppm different VOCs at 200 °C, (<b>e</b>) the response of ZnO/In<sub>2</sub>O<sub>3</sub> sensor to 100 ppm mixed VOCs at 200 °C.</p> Full article ">
14 pages, 3833 KiB  
Article
Real-Time Indoor Visible Light Positioning (VLP) Using Long Short Term Memory Neural Network (LSTM-NN) with Principal Component Analysis (PCA)
by Yueh-Han Shu, Yun-Han Chang, Yuan-Zeng Lin and Chi-Wai Chow
Sensors 2024, 24(16), 5424; https://doi.org/10.3390/s24165424 - 22 Aug 2024
Cited by 1 | Viewed by 569
Abstract
New applications such as augmented reality/virtual reality (AR/VR), Internet-of-Things (IOT), autonomous mobile robot (AMR) services, etc., require high reliability and high accuracy real-time positioning and tracking of persons and devices in indoor areas. Among the different visible-light-positioning (VLP) schemes, such as proximity, time-of-arrival [...] Read more.
New applications such as augmented reality/virtual reality (AR/VR), Internet-of-Things (IOT), autonomous mobile robot (AMR) services, etc., require high reliability and high accuracy real-time positioning and tracking of persons and devices in indoor areas. Among the different visible-light-positioning (VLP) schemes, such as proximity, time-of-arrival (TOA), time-difference-of-arrival (TDOA), angle-of-arrival (AOA), and received-signal-strength (RSS), the RSS scheme is relatively easy to implement. Among these VLP methods, the RSS method is simple and efficient. As the received optical power has an inverse relationship with the distance between the LED transmitter (Tx) and the photodiode (PD) receiver (Rx), position information can be estimated by studying the received optical power from different Txs. In this work, we propose and experimentally demonstrate a real-time VLP system utilizing long short-term memory neural network (LSTM-NN) with principal component analysis (PCA) to mitigate high positioning error, particularly at the positioning unit cell boundaries. Experimental results show that in a positioning unit cell of 100 × 100 × 250 cm3, the average positioning error is 5.912 cm when using LSTM-NN only. By utilizing the PCA, we can observe that the positioning accuracy can be significantly enhanced to 1.806 cm, particularly at the unit cell boundaries and cell corners, showing a positioning error reduction of 69.45%. In the cumulative distribution function (CDF) measurements, when using only the LSTM-NN model, the positioning error of 95% of the experimental data is >15 cm; while using the LSTM-NN with PCA model, the error is reduced to <5 cm. In addition, we also experimentally demonstrate that the proposed real-time VLP system can also be used to predict the direction and the trajectory of the moving Rx. Full article
(This article belongs to the Special Issue Challenges and Future Trends in Optical Communications)
►▼ Show Figures

Figure 1

Figure 1
<p>(<b>a</b>) Architecture of the VLP system with four LEDs modulated by specific RF carrier frequencies of <span class="html-italic">f</span><sub>1</sub>, <span class="html-italic">f</span><sub>2</sub>, <span class="html-italic">f</span><sub>3</sub>, and <span class="html-italic">f</span><sub>4</sub>, (47 kHz, 59 kHz, 83 kHz, 101 kHz), respectively. (<b>b</b>) Bird-view of the positioning unit cell indicating the training and testing locations.</p> Full article ">Figure 2
<p>(<b>a</b>) Experimental photo of the VLP experiment. (<b>b</b>) Photo of the client side. The PD, RTO, and PC are all placed on a trolley for training and testing data collections. PD: photodiode; RTO: real-time oscilloscope.</p> Full article ">Figure 3
<p>Architecture of the VLP Rx. ID: optical identifier; BPF: band-pass filter; LPF: low-pass filter.</p> Full article ">Figure 4
<p>Flow diagram of the proposed real-time VLP system utilizing LSTM-NN with PCA.</p> Full article ">Figure 5
<p>Flow diagram of the PCA used in the VLP experiment.</p> Full article ">Figure 6
<p>Structure of an LSTM cell used in the LSTM-NN model.</p> Full article ">Figure 7
<p>Structure of the proposed LSTM-NN model used in both training phase and testing phase.</p> Full article ">Figure 8
<p>Error distributions using (<b>a</b>) the LSTM-NN only and (<b>b</b>) the LSTM-NN with PCA.</p> Full article ">Figure 9
<p>CDF of the measured positioning error using LSTM-NN only and using the LSTM-NN with PCA.</p> Full article ">Figure 10
<p>Error distributions using (<b>a</b>) FCN only and (<b>b</b>) FCN with PCA.</p> Full article ">Figure 11
<p>CDF of the measured positioning error using FCN only and using FCN with PCA.</p> Full article ">Figure 12
<p>Experimental predicted location of the moving Rx using the LSTM-NN with PCA at different iterations. (<b>a</b>–<b>h</b>) Indication of predicted direction and trajectory of the Rx from iteration 1 to 7.</p> Full article ">
20 pages, 4847 KiB  
Article
A Small-Object-Detection Algorithm Based on LiDAR Point-Cloud Clustering for Autonomous Vehicles
by Zhibing Duan, Jinju Shao, Meng Zhang, Jinlei Zhang and Zhipeng Zhai
Sensors 2024, 24(16), 5423; https://doi.org/10.3390/s24165423 - 22 Aug 2024
Viewed by 952
Abstract
3D object-detection based on LiDAR point clouds can help driverless vehicles detect obstacles. However, the existing point-cloud-based object-detection methods are generally ineffective in detecting small objects such as pedestrians and cyclists. Therefore, a small-object-detection algorithm based on clustering is proposed. Firstly, a new [...] Read more.
3D object-detection based on LiDAR point clouds can help driverless vehicles detect obstacles. However, the existing point-cloud-based object-detection methods are generally ineffective in detecting small objects such as pedestrians and cyclists. Therefore, a small-object-detection algorithm based on clustering is proposed. Firstly, a new segmented ground-point clouds segmentation algorithm is proposed, which filters out the object point clouds according to the heuristic rules and realizes the ground segmentation by multi-region plane-fitting. Then, the small-object point cloud is clustered using an improved DBSCAN clustering algorithm. The K-means++ algorithm for pre-clustering is used, the neighborhood radius is adaptively adjusted according to the distance, and the core point search method of the original algorithm is improved. Finally, the detection of small objects is completed using the directional wraparound box model. After extensive experiments, it was shown that the precision and recall of our proposed ground-segmentation algorithm reached 91.86% and 92.70%, respectively, and the improved DBSCAN clustering algorithm improved the recall of pedestrians and cyclists by 15.89% and 9.50%, respectively. In addition, visualization experiments confirmed that our proposed small-object-detection algorithm based on the point-cloud clustering method can realize the accurate detection of small objects. Full article
(This article belongs to the Section Vehicular Sensing)
►▼ Show Figures

Graphical abstract

Graphical abstract
Full article ">Figure 1
<p>Improved small-object-detection algorithm flow based on the point-cloud clustering method.</p> Full article ">Figure 2
<p>Segmented ground-point cloud accurate segmentation algorithm flow chart.</p> Full article ">Figure 3
<p>Example of object and ground classification. (<b>a</b>–<b>c</b>), respectively, describe the situations of object points and ground points under different road conditions.</p> Full article ">Figure 4
<p>Multi-region polar gridding. The blue line is the area division line.</p> Full article ">Figure 5
<p>Seed-point set iteration process diagram. (<b>a</b>–<b>c</b>) depict the first, second, and nth iterations of the seed-point set, respectively. The red points are ground points.</p> Full article ">Figure 6
<p>Schematic diagram of the reference normal vector, with the referenced subgrid in red and its surrounding neighborhood grid in green.</p> Full article ">Figure 7
<p>Improved way of searching for core points.</p> Full article ">Figure 8
<p>Comparison of segmentation results in the rough road scene. The first row corresponds to the results processed by the RANSAC algorithm, the second row corresponds to the results of the R-GPF algorithm, and the third row corresponds to the results of our proposed ground-segmentation algorithm. The left side (<b>a</b>) is the scene map captured by the front-view camera, the center position (<b>b</b>,<b>d</b>,<b>f</b>) is the global ground-segmentation effect map of each algorithm, and the right side (<b>c</b>,<b>e</b>,<b>g</b>) is the local ground-segmentation effect map of each algorithm, in which the green dots are the ground dots and the red dots are the non-ground dots.</p> Full article ">Figure 9
<p>Comparison of segmentation results for the sloped pavement scene. The first row corresponds to the results processed by the RANSAC algorithm, the second row corresponds to the results of the R-GPF algorithm, and the third row corresponds to the results of our proposed ground-segmentation algorithm. The left side (<b>a</b>) is the scene map captured by the front-view camera, the center position (<b>b</b>,<b>d</b>,<b>f</b>) is the global ground-segmentation effect map of each algorithm, and the right side (<b>c</b>,<b>e</b>,<b>g</b>) is the local ground-segmentation effect map of each algorithm, in which the green dots are the ground dots and the red dots are the non-ground dots.</p> Full article ">Figure 10
<p>Comparison of the visualization results of the two algorithms for clustering detection. The left side (<b>a</b>,<b>d</b>) shows different scene maps captured by the camera, the middle (<b>b</b>,<b>e</b>) shows the clustering detection results of the traditional DBSCAN algorithm, and the right side (<b>c</b>,<b>f</b>) shows the clustering detection results of our algorithm. The blue points in the detection map are point clouds, and the green box is the bounding box fitted by the algorithm.</p> Full article ">Figure 11
<p>Comparison of two algorithms for different types of clustering detection results.</p> Full article ">
23 pages, 11964 KiB  
Article
Research on Corn Leaf and Stalk Recognition and Ranging Technology Based on LiDAR and Camera Fusion
by Xueting Hu, Xiao Zhang, Xi Chen and Lu Zheng
Sensors 2024, 24(16), 5422; https://doi.org/10.3390/s24165422 - 22 Aug 2024
Viewed by 504
Abstract
Corn, as one of the three major grain crops in China, plays a crucial role in ensuring national food security through its yield and quality. With the advancement of agricultural intelligence, agricultural robot technology has gained significant attention. High-precision navigation is the basis [...] Read more.
Corn, as one of the three major grain crops in China, plays a crucial role in ensuring national food security through its yield and quality. With the advancement of agricultural intelligence, agricultural robot technology has gained significant attention. High-precision navigation is the basis for realizing various operations of agricultural robots in corn fields and is closely related to the quality of operations. Corn leaf and stalk recognition and ranging are the prerequisites for achieving high-precision navigation and have attracted much attention. This paper proposes a corn leaf and stalk recognition and ranging algorithm based on multi-sensor fusion. First, YOLOv8 is used to identify corn leaves and stalks. Considering the large differences in leaf morphology and the large changes in field illumination that lead to discontinuous identification, an equidistant expansion polygon algorithm is proposed to post-process the leaves, thereby increasing the average recognition completeness of the leaves to 86.4%. Secondly, after eliminating redundant point clouds, the IMU data are used to calculate the confidence of the LiDAR and depth camera ranging point clouds, and point cloud fusion is performed based on this to achieve high-precision ranging of corn leaves. The average ranging error is 2.9 cm, which is lower than the measurement error of a single sensor. Finally, the stalk point cloud is processed and clustered using the FILL-DBSCAN algorithm to identify and measure the distance of the same corn stalk. The algorithm combines recognition accuracy and ranging accuracy to meet the needs of robot navigation or phenotypic measurement in corn fields, ensuring the stable and efficient operation of the robot in the corn field. Full article
(This article belongs to the Section Smart Agriculture)
►▼ Show Figures

Figure 1

Figure 1
<p>Flowchart of the algorithm.</p> Full article ">Figure 2
<p>Annotations of Corn Leaves and Stalks in the YOLOv8 Model Training Dataset. Green outline indicates corn leaves; Purple outline indicates corn stalks.</p> Full article ">Figure 3
<p>Process of equidistant contour generation: (<b>a</b>) actual shape of the leaf, (<b>b</b>) leaf Recognition results under YOLOv8, (<b>c</b>) drawing of equidistant circles around leaf contour points, (<b>d</b>) fitting of equidistant contours around leaf.</p> Full article ">Figure 4
<p>Post-processing results of leaf recognition using equidistant contour clustering. Red, green, and blue areas represent the identified discontinuous leaves. The red contour line shows the complete leaf outline after clustering using the equidistant expansion polygon post-processing algorithm.</p> Full article ">Figure 5
<p>Misidentification of stems under leaf occlusion. (<b>a</b>) Real scene of corn stalk testing. (<b>b</b>) DBSCAN clustering of detected stalk occlusion point clouds. (<b>c</b>) Due to leaf overlap, a single corn stalk is clustered into multiple point cloud sets.</p> Full article ">Figure 6
<p>Post-processing of stem with equidistant outward-expanded polygon.</p> Full article ">Figure 7
<p>Definition of the XYZ directions for the corn stalk.</p> Full article ">Figure 8
<p>Training results of YOLOv8l model.</p> Full article ">Figure 9
<p>Simulation of curved leaf morphology.</p> Full article ">Figure 10
<p>Test of equidistant expansion polygon post-processing algorithm: (<b>a</b>) Recognition of uniform leaf morphology. (<b>b</b>) Recognition of curved leaf morphology. (<b>c</b>) Result of leaf post-processing with equidistant outward-expanded polygons.</p> Full article ">Figure 10 Cont.
<p>Test of equidistant expansion polygon post-processing algorithm: (<b>a</b>) Recognition of uniform leaf morphology. (<b>b</b>) Recognition of curved leaf morphology. (<b>c</b>) Result of leaf post-processing with equidistant outward-expanded polygons.</p> Full article ">Figure 11
<p>Test of leaf post-processing with equidistant outward-expanded polygons. ICP: identification completeness percentage.</p> Full article ">Figure 12
<p>Laser displacement sensor test platform.</p> Full article ">Figure 13
<p>Laser displacement sensor measurement test setup: (<b>a</b>) Frontal scanning measurement; (<b>b</b>) Rear scanning measurement.</p> Full article ">Figure 14
<p>Leaf scanning results: (<b>a</b>) distance point cloud obtained from scanning; (<b>b</b>) leaf convex hull shape extracted using the α-shape algorithm.</p> Full article ">Figure 15
<p>Measurement point setup.</p> Full article ">Figure 16
<p>Experimental prototype.</p> Full article ">Figure 17
<p>Upper interface of prototype test: (<b>a</b>) shows the physical maize leaves and recognition results, (<b>b</b>) displays the distance measurement results from the depth camera and LiDAR point clouds.</p> Full article ">Figure 18
<p>Interface of the prototype test on the host computer: (<b>a</b>) Real corn leaf and recognition results. The blue outline represents the identified contours, and the red area indicates the identified mask. (<b>b</b>) Depth camera and LiDAR point cloud measurement results.</p> Full article ">Figure 19
<p>Stalk FILL-DBSCAN test: (<b>a</b>) corn model, (<b>b</b>) segmented corn stems identified by YOLOv8, (<b>c</b>) plane clustering using DBSCAN algorithm after point cloud projection, (<b>d</b>) three-dimensional clustering of segmented point clouds of the same corn stem.</p> Full article ">Figure 20
<p>Actual conditions of the testing scenario in the corn field.</p> Full article ">Figure 21
<p>Field test schematic. (<b>a</b>) Schematic diagram of measurement point 1; (<b>b</b>) Schematic diagram of measurement point 2.</p> Full article ">Figure 22
<p>The interface of the host computer Rviz during the test.</p> Full article ">
Show export options expand_more Show export options expand_less
Select all
Export citation of selected articles as:
 
Displaying articles 1-371
Previous Issue
Next Issue
Back to TopTop