CN109341703A - A full-cycle visual SLAM algorithm using CNNs feature detection - Google Patents
A full-cycle visual SLAM algorithm using CNNs feature detection Download PDFInfo
- Publication number
- CN109341703A CN109341703A CN201811087509.9A CN201811087509A CN109341703A CN 109341703 A CN109341703 A CN 109341703A CN 201811087509 A CN201811087509 A CN 201811087509A CN 109341703 A CN109341703 A CN 109341703A
- Authority
- CN
- China
- Prior art keywords
- training
- feature
- cnns
- detection
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01C—MEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
- G01C21/00—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00
- G01C21/26—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00 specially adapted for navigation in a road network
- G01C21/28—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00 specially adapted for navigation in a road network with correlation of data from several navigational instruments
- G01C21/30—Map- or contour-matching
- G01C21/32—Structuring or formatting of map data
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Remote Sensing (AREA)
- Radar, Positioning & Navigation (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Automation & Control Theory (AREA)
- Image Analysis (AREA)
Abstract
本发明公开一种全周期采用CNNs特征检测的视觉SLAM算法，首先在前端，用无监督模型对原始图像数据进行预训练，接着利用预训练的数据通过CNN网络架构把运动和深度的联合表示和局部速度和方向改变关联起来，执行视觉里程计；最后执行路径预测。本发明还使用OverFeat神经网络模型进行回环检测环节，用于消除前端带来的累积误差，构建一个基于深度学习的视觉slam架构。同时构建时间和空间连续性滤波器，对匹配结果进行验证，提高匹配准确率，消除误匹配。本发明在提高视觉里程计精度和闭环检测正确率方面有着巨大的优势和潜力。
The invention discloses a visual SLAM algorithm that adopts CNNs feature detection in the whole cycle. First, at the front end, the original image data is pre-trained with an unsupervised model, and then the pre-trained data is used to combine the joint representation of motion and depth with the CNN network architecture. Local velocity and direction changes are correlated to perform visual odometry; finally, path prediction is performed. The present invention also uses the OverFeat neural network model to perform the loopback detection link, which is used to eliminate the accumulated error brought by the front end and construct a visual slam architecture based on deep learning. At the same time, time and space continuity filters are constructed to verify the matching results, improve the matching accuracy and eliminate false matching. The present invention has great advantages and potential in improving the accuracy of the visual odometry and the correct rate of closed-loop detection.
Description
Technical field
The technical field with map structuring (slam) algorithm is positioned while the invention belongs in computer vision, is specifically related to
And a kind of complete period uses the vision SLAM algorithm of CNNs feature detection.
Background technique
SLAM (Simultaneous Localization and Mapping) Chinese be " while positioning and map
Building ".SLAM is a fascinating research field, has in robot, navigation and other many applications and widely answers
With.Vision SLAM is essentially related to estimate camera motion according to visual sensor information and is attempted to construct the ground of ambient enviroment
Figure, such as the sequence frame from one or more cameras.The research means of current SLAM problem are mainly by robot sheet
Sensor is installed to estimate the characteristic information of robot body motion information and circumstances not known on body, is merged using information
It realizes to the accurate estimation of robot pose and the spatial modeling of scene.Although the sensor that SLAM is used has laser and vision
Equal multiple types, but its treatment process generally comprises 3 parts: front-end vision odometer, rear end optimization and closed loop detection.
Typical vision SLAM algorithm is to estimate pose of camera as main target, 3D is reconstructed by multiple view geometry theory
Figure.To improve data processing speed, partial visual SLAM algorithm extracts sparse characteristics of image first, by between characteristic point
Visual odometry and closed loop detection are realized in matching, such as special based on SIFT (scale invariant feature transform)
The vision SLAM [13] of the sign and vision SLAM for being based on ORB (oriented FAST and rotated BRIEF) feature.SIFT
Rely on its preferable robustness and preferably separating capacity and quick processing speed with ORB feature, in the field vision SLAM
It is widely applied.The sparse characteristics of image of engineer currently has many limitations, and it is special on the one hand how to design sparse image
Sign optimally indicates that image information is still the unsolved major issue of computer vision field, on the other hand sparse characteristics of image
In terms of reply illumination variation, dynamic object movement, camera parameters change and lack texture or texture is single
Still there is more challenge.Traditional visual odometry (Visual Odometry, VO) essentially relates to be estimated according to visual information
Meter movement, such as the sequence frame from one or more cameras.One common feature of most number attribute in these work is
They estimate visual token by critical point detection and tracking and camera geometry.
In recent years, the method based on study many fields of computer vision show desirable as a result,
Defect present in traditional vision slam algorithm can be overcome, and (sparse characteristics of image is in reply illumination variation, dynamic object fortune
Dynamic, camera parameters, which change and lack texture or the single environment etc. of texture, more difficulty).As convolutional Neural net
Model as network (CNNs) is proved to highly effective in various visual tasks, such as classifies and positions, estimation of Depth etc..Nothing
Supervise the ability that feature learning model display is indicated by partial transformation in multiplication interactive learning data.Research shows that by nothing
Data after monitor model pre-training reapply in CNNs network, preferable can cross noise filtering and prevent " over-fitting ".
Visual odometry (VO) is responsible for the initial value of estimation track and map.VO principle is the connection that only considered consecutive frame picture
System.Error inevitably will be with time integral, so that cumulative errors occurs in whole system, the result estimated for a long time will not
Reliably, in other words, we can not construct globally consistent track and map.Winding detection module can provide one except consecutive frame
A bit after constraint more remote.The key of winding detection, be how effectively to detect camera by the same place this
Part thing is related to the correctness of the estimation and map of our estimations under long-time.Therefore, winding detection is to entire SLAM system
The promotion of precision and robustness be obviously.
And this winding detection based on appearance is substantially a task of image similarity match.By scheming to two
As carrying out characteristic matching, to verify whether as same place.The conventional method of winding detection is that " bag of words " is utilized to generate word
Allusion quotation does characteristic matching.Now, CNNs shows optimal performance in various classification tasks.Existing road sign test result
Show that the further feature of the CNN from different levels executes more preferably than SIFT always in terms of descriptors match, shows SIFT
Or SURF may no longer be the preferred descriptor of matching task.Therefore, our invention by CNN in terms of image classification go out
The inspiration for the feasibility evidence that color table is existing and they are in terms of characteristic matching.The bag of words method to discard tradition, with being based on
CNNs using depth learning technology as the stratification image characteristic extracting method of representative, to do winding detection.Deep learning algorithm
It is the recognizer of computer vision field mainstream, relies on the stratification mark sheet of multilayer neural network study image
Show, compared with traditional recognition method, the accuracy rate of higher feature extraction and position identification may be implemented.
Summary of the invention
In view of the above-mentioned problems, the present invention proposes that a kind of complete period uses the vision SLAM algorithm of CNNs feature detection, it is one
Kind is all handled using convolutional neural networks in the front end (VO) that SLAM algorithm is run and winding detection, realizes a complete period
The SLAM system of deep learning algorithm.
Complete period of the invention uses the vision SLAM algorithm of CNNs feature detection, comprising the following steps:
Step 1: using the environmental information around binocular camera scanning；A part in collected multiple video flowings is made
For training dataset, a part is used as test data set.
Step 2: training dataset carries out pre-training in the video flowing acquired by the method for synchronous detection to step 1.
Step 3: being changed using the part that convolutional neural networks training obtains speed and changed with the part in direction, to execute view
Feel odometer.
Step 4: the part of the speed obtained using step 3, which is changed, changes information with the part in direction, recovers camera
Motion path.
Step 5: carrying out closed loop detection using convolutional neural networks, eliminate the accumulated error of path prediction.
The invention has the advantages that:
1, the complete period of the invention uses the vision SLAM algorithm of CNNs feature detection, and convolutional neural networks are used for front end,
Compared to conventional front-end algorithm, the side based on study substitutes cumbersome formula and calculates, and extracts and matches without manual features, it appears succinct
Intuitively, and on-line operation fast speed.
2, the complete period of the invention uses the vision SLAM algorithm of CNNs feature detection, the utilization bag of words to discard tradition into
The method of row closed loop detection obtains the effect of relatively good position identification by accurate characteristic matching.
3, the complete period of the invention uses the vision SLAM algorithm of CNNs feature detection, can pass through neural network learning image
In profound feature, discrimination can achieve higher level.Compared to Conventional visual slam algorithm, closed loop inspection can be improved
Survey accuracy rate, expression image information it is more abundant, have more strong robustness to environmental changes such as illumination, seasons, can to 2 frame images it
Between similitude calculated, to realize more succinct visual odometry, and complete using database pre-training neural network
While at characteristic Design, the design of matched classifier can also be synchronously completed.
Detailed description of the invention
Fig. 1 is the vision SLAM algorithm overall flow figure for using the detection of CNNs feature in the complete period of the invention；
Fig. 2 is the complete period of the invention using closed loop detection method flow chart in the vision SLAM algorithm of CNNs feature detection.
Specific embodiment
The present invention is described in further detail below in conjunction with the accompanying drawings.
Complete period of the invention uses the vision SLAM algorithm of CNNs feature detection, as shown in Figure 1, comprising the following steps:
Step 1: scanning the environmental information of surrounding；
Using binocular camera along being moved in square area, ambient image information acquisition is carried out to real scene, is obtained
Video flowing be transmitted in host computer in real time.Binocular camera mobile ring number is 1~2 circle, constitutes closed loop, facilitates subsequent close
Compensation of the ring detection to cumulative errors.Repeat the above process, by a part of conduct in collected multiple video flowings
Training dataset, a part are used as test data set.
Step 2: training dataset carries out pre-training in the video flowing acquired by the method for synchronous detection to step 1.
It indicates to obtain the combination of camera motion and image depth information, is adopted using unsupervised learning model (SAE-D)
Pre-training is carried out to training dataset with stochastic gradient descent training.It is the model of a single layer based on synchronous SAE-D, allows
Feature is obtained from training dataset by part and Hebbian type study.
SAE-D model concentrates training, the size of each binocular block on the local binocular block cut out immediately in training data
For 16*16*5 (space*space*time), the characteristic information of movement and depth joint expression is obtained；It then will movement and depth
The characteristic information that joint indicates goes albefaction back to image space, completes the pre-training to training dataset.Instruction after pre-training
Practice the first layer that data set is used to initialize CNN (convolutional neural networks).
Step 3: executing visual odometry using convolutional neural networks (CNN) training.
Convolutional neural networks (CNN) are a kind of learning models based on supervision.CNN is trained to by partial-depth and movement
Indicate associated with the localized variation in speed and direction, to learn to execute visual odometry.By CNN framework the fortune of acquisition
Dynamic and depth joint expression is trained associated with desired label (direction and speed change).
The feature that SAE-D model training obtains in step 2 is inputted to the CNN first layer of two same architectures in the present invention,
Initialization for two CNN first layers；The part of two CNN difference output speeds changes to be changed with the part in direction, and is had
It is associated with desired label that the part of speed is changed the part change with direction respectively by two CNN remainders.
Entire CNN network shares 6 layers, and first layer is the convolutional layer of 5*5, respectively left images learning characteristic.Following
Two layers of feature extracted to two convolutional layers in left and right carry out element multiplication.The convolutional layer of third layer 1*1, the 4th layer is pond
Layer, layer 5 is full articulamentum, and last output layer is Softmax layers.
The input of above-mentioned two CNN is 5 frame subsequences, and target output is local velocity and the vector that direction changes
It indicates.The speed of part and direction, which change the effect of information and precision, to be evaluated and tested by binocular data set KITTI.
Step 4: the prediction of camera motion path
For entire video flowing, the speed of the every 5 frame subsequence obtained using step 3 and direction change information, Ji Keli
The scattered motion path for recovering camera.
Step 5: carrying out closed loop detection using CNN, eliminate the accumulated error of path prediction；
Local speed obtained in step 4 and direction change information can not completely precisely, and there is a certain error.With
The continuous accumulation of error, predicted path can constantly increase with the gap in true path from origin-to-destination.So needing
Subsequent closed loop detection eliminates these cumulative errors, reduces the gap of predicted path and true path.It is this
Algorithm includes two parts: carrying out feature extraction using convolutional neural networks；Location matches are assumed by comparing characteristic response
Carry out spatio-temporal filtering.
Equally handled using the algorithm based on CNN in closed loop detection, the CNN model being applied to it is aforementioned
CNN model is different, is independently used in visual odometry and closed loop detection, pre- with the camera motion path that this removal process 4 obtains
The cumulative errors of survey realize from main closed loop, method particularly includes:
Image characteristics extraction is carried out using the convolutional neural networks of pre-training.The present invention uses overfeat convolutional Neural net
Network carries out image characteristics extraction.
Overfeat convolutional neural networks carry out pre-training on 2012 data set of ImageNet, and the data set is by 1,200,000
Open image and 1000 classes compositions.Overfeat convolutional neural networks include that five convolution stages and three are fully connected the stage.
Two convolution stages of beginning are by a convolutional layer, a maximum pond layer and rectification (ReLU) non-linear layer composition.The
Three and the Volume Four product stage by convolutional layer, zero padding layer and ReLU non-linear layer composition.5th stage include a convolutional layer, one
A zero padding layer, one ReLU layers and one Maxpooling layers.Finally, the 6th and the 7th be fully connected the stage include one it is complete
Full articulamentum and one ReLU layer, and it includes the output layer for being only fully connected layer that the 8th stage, which was,.Entire convolutional neural networks are total
Share 21 layers.
When image I is input into network, it generates a series of layering activation.L is used in the present inventionk(I), k=
1 ..., 21 indicate the corresponding output of given input picture I kth layer.The feature vector L of each layer of outputk(I) every in
One be all image I deep learning indicate；Execution position is compared to the corresponding feature vector of different images by these
Set identification.The network is capable of handling the image that any size is equal to or more than 231 × 231 pixels, therefore overfeat convolution mind
Input through network is all using being sized as the image of 256 × 256 pixels.
It is good using pre-training as a result, by using the training dataset acquired in step 1 and test data set as input
Overfeat convolutional neural networks carry out feature extraction.
Step 6: characteristic matching generates hybrid matrix and carries out space-time expending detection.
As shown in Fig. 2, being extracted by overfeat convolutional neural networks to from the picture that each test data is concentrated
Feature concentrate the feature proposed to be matched with from each training data.
To compare performance difference of the feature of every tomographic image in overfeat convolutional neural networks in scene Recognition, into one
Step utilizes each layer latent structure hybrid matrix:
Mk(i, j)=d (Lk(Ii),Lk(Ij)), i=1 ..., R, j=1 ..., T
Wherein, IiRepresent the image input of the i-th frame training data concentration, IjThe image for representing jth frame test data concentration is defeated
Enter, Lk(Ii) represent and IiCorresponding kth layer output, Mk(i, j) represents the Europe between kth layer training sample i and test sample j
Family name's distance describes matching degree between the two；R and T respectively represent the number of training image and the number of test image.
The averaged feature vector that the above-mentioned each column of hybrid matrix store between jth frame test image and all training images is poor.
In order to find strongest location matches it is assumed that having the member of minimum feature vector difference in search hybrid matrix each column
Element.
For location matches possible in hybrid matrix it is assumed that further constructing spatial continuity filter and Time Continuous
Property filter carry out comprehensive verification, improve matching accuracy rate.Meanwhile the characteristic performance come is trained to each layer network and is visited
Rope, it is preferable that discovery network middle layer feature describes similar for visual angle image matching effect, and in after layer for scene visual angle
Variation has stronger adaptability and robustness.
There are accurately location matches, so that it may the cumulative errors caused by the visual odometry of not winding compensate,
Construct globally consistent track.
Claims (5)
1. the vision SLAM algorithm that a kind of complete period uses the detection of CNNs feature, comprising the following steps:
Step 1: using the environmental information around binocular camera scanning；By a part in collected multiple video flowings as instruction
Practice data set, a part is used as test data set；
Step 2: training dataset carries out pre-training in the video flowing acquired by the method for synchronous detection to step 1；
Step 3: being changed using the part that convolutional neural networks training obtains speed and changed with the part in direction, to execute in vision
Journey meter；
Step 4: the part of the speed obtained using step 3, which is changed, changes information with the part in direction, recovers the movement of camera
Path；
Step 5: carrying out closed loop detection using convolutional neural networks, eliminate the accumulated error of path prediction.
2. a kind of complete period uses the vision SLAM algorithm of CNNs feature detection as described in claim 1, it is characterised in that: step
Using binocular camera along moving in annular region in 1, mobile ring number is 1~2 circle, constitutes closed loop.
3. a kind of complete period uses the vision SLAM algorithm of CNNs feature detection as described in claim 1, it is characterised in that: step
Pre-training is carried out to training dataset using stochastic gradient descent training using unsupervised learning model in 2, is moved and deep
The characteristic information that movement and depth joint indicate then is gone albefaction back to image space by the characteristic information that degree joint indicates.
4. a kind of complete period uses the vision SLAM algorithm of CNNs feature detection as described in claim 1, it is characterised in that: step
In 3, the feature that step 2 training obtains is inputted to the CNN first layer of two same architectures, by CNN framework the movement of acquisition
Joint expression with depth is trained associated with desired label.
5. a kind of complete period uses the vision SLAM algorithm of CNNs feature detection as described in claim 1, it is characterised in that: step
In 5, firstly, by using the training dataset acquired in step 1 and test data set as input, using in imagenet number
According to upper pre-training is collected, good overfeat convolutional neural networks carry out feature extraction；Then, it is concentrated to from each test data
Picture in the feature extracted concentrate the feature proposed to be matched with from each training data；
Utilize each layer latent structure hybrid matrix of overfeat convolutional neural networks:
Mk(i, j)=d (Lk(Ii),Lk(Ij)), i=1 ..., R, j=1 ..., T
Wherein, IiRepresent the image input of the i-th frame training data concentration, IjThe image input of jth frame test data concentration is represented,
Lk(Ii) represent and IiCorresponding kth layer output, Mk(i, j) represent the Euclidean between kth layer training sample i and test sample j away from
From describing matching degree between the two；R and T respectively represent the number of training image and the number of test image；It is above-mentioned
The averaged feature vector that each column of hybrid matrix store between jth frame test image and all training images is poor；
Search for the element in hybrid matrix each column with minimum feature vector difference；
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811087509.9A CN109341703B (en) | 2018-09-18 | 2018-09-18 | Visual SLAM algorithm adopting CNNs characteristic detection in full period |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811087509.9A CN109341703B (en) | 2018-09-18 | 2018-09-18 | Visual SLAM algorithm adopting CNNs characteristic detection in full period |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109341703A true CN109341703A (en) | 2019-02-15 |
| CN109341703B CN109341703B (en) | 2022-07-01 |
Family
ID=65305452
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811087509.9A Active CN109341703B (en) | 2018-09-18 | 2018-09-18 | Visual SLAM algorithm adopting CNNs characteristic detection in full period |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109341703B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109840598A (en) * | 2019-04-29 | 2019-06-04 | 深兰人工智能芯片研究院（江苏）有限公司 | A kind of method for building up and device of deep learning network model |
| CN110146099A (en) * | 2019-05-31 | 2019-08-20 | 西安工程大学 | A Synchronous Localization and Map Construction Method Based on Deep Learning |
| CN110296705A (en) * | 2019-06-28 | 2019-10-01 | 苏州瑞久智能科技有限公司 | A kind of vision SLAM winding detection method based on learning distance metric |
| CN110487274A (en) * | 2019-07-30 | 2019-11-22 | 中国科学院空间应用工程与技术中心 | SLAM method, system, navigation vehicle and storage medium for weak texture scene |
| CN110659619A (en) * | 2019-09-27 | 2020-01-07 | 昆明理工大学 | Depth space-time information-based correlation filtering tracking method |
| CN110738128A (en) * | 2019-09-19 | 2020-01-31 | 天津大学 | repeated video detection method based on deep learning |
| CN111144550A (en) * | 2019-12-27 | 2020-05-12 | 中国科学院半导体研究所 | Simplex Deep Neural Network Model and Construction Method Based on Homologous Continuity |
| CN111241986A (en) * | 2020-01-08 | 2020-06-05 | 电子科技大学 | A closed-loop detection method for visual SLAM based on end-to-end relational network |
| CN111243021A (en) * | 2020-01-06 | 2020-06-05 | 武汉理工大学 | Vehicle-mounted visual positioning method and system based on multiple combined cameras and storage medium |
| CN111753789A (en) * | 2020-07-01 | 2020-10-09 | 重庆邮电大学 | Robot Vision SLAM Closed-loop Detection Method Based on Stacked Combined Autoencoders |
| CN113066152A (en) * | 2021-03-18 | 2021-07-02 | 内蒙古工业大学 | AGV map construction method and system |
| CN110399821B (en) * | 2019-07-17 | 2023-05-30 | 上海师范大学 | Customer satisfaction acquisition method based on facial expression recognition |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106651830A (en) * | 2016-09-28 | 2017-05-10 | 华南理工大学 | Image quality test method based on parallel convolutional neural network |
| CN106780631A (en) * | 2017-01-11 | 2017-05-31 | 山东大学 | A kind of robot closed loop detection method based on deep learning |
| CN107563430A (en) * | 2017-08-28 | 2018-01-09 | 昆明理工大学 | A kind of convolutional neural networks algorithm optimization method based on sparse autocoder and gray scale correlation fractal dimension |
| CN107808132A (en) * | 2017-10-23 | 2018-03-16 | 重庆邮电大学 | A kind of scene image classification method for merging topic model |
| CN107944386A (en) * | 2017-11-22 | 2018-04-20 | 天津大学 | Visual scene recognition methods based on convolutional neural networks |
-
2018
- 2018-09-18 CN CN201811087509.9A patent/CN109341703B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106651830A (en) * | 2016-09-28 | 2017-05-10 | 华南理工大学 | Image quality test method based on parallel convolutional neural network |
| CN106780631A (en) * | 2017-01-11 | 2017-05-31 | 山东大学 | A kind of robot closed loop detection method based on deep learning |
| CN107563430A (en) * | 2017-08-28 | 2018-01-09 | 昆明理工大学 | A kind of convolutional neural networks algorithm optimization method based on sparse autocoder and gray scale correlation fractal dimension |
| CN107808132A (en) * | 2017-10-23 | 2018-03-16 | 重庆邮电大学 | A kind of scene image classification method for merging topic model |
| CN107944386A (en) * | 2017-11-22 | 2018-04-20 | 天津大学 | Visual scene recognition methods based on convolutional neural networks |
Non-Patent Citations (1)
| Title |
|---|
| 何元烈等: "基于精简卷积神经网络的快速闭环检测方法", 《计算机工程》 * |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109840598B (en) * | 2019-04-29 | 2019-08-09 | 深兰人工智能芯片研究院（江苏）有限公司 | A method and device for establishing a deep learning network model |
| CN109840598A (en) * | 2019-04-29 | 2019-06-04 | 深兰人工智能芯片研究院（江苏）有限公司 | A kind of method for building up and device of deep learning network model |
| CN110146099A (en) * | 2019-05-31 | 2019-08-20 | 西安工程大学 | A Synchronous Localization and Map Construction Method Based on Deep Learning |
| CN110296705B (en) * | 2019-06-28 | 2022-01-25 | 苏州瑞久智能科技有限公司 | Visual SLAM loop detection method based on distance metric learning |
| CN110296705A (en) * | 2019-06-28 | 2019-10-01 | 苏州瑞久智能科技有限公司 | A kind of vision SLAM winding detection method based on learning distance metric |
| CN110399821B (en) * | 2019-07-17 | 2023-05-30 | 上海师范大学 | Customer satisfaction acquisition method based on facial expression recognition |
| CN110487274A (en) * | 2019-07-30 | 2019-11-22 | 中国科学院空间应用工程与技术中心 | SLAM method, system, navigation vehicle and storage medium for weak texture scene |
| CN110487274B (en) * | 2019-07-30 | 2021-01-29 | 中国科学院空间应用工程与技术中心 | SLAM method and system for weak texture scene, navigation vehicle and storage medium |
| CN110738128A (en) * | 2019-09-19 | 2020-01-31 | 天津大学 | repeated video detection method based on deep learning |
| CN110659619A (en) * | 2019-09-27 | 2020-01-07 | 昆明理工大学 | Depth space-time information-based correlation filtering tracking method |
| CN111144550A (en) * | 2019-12-27 | 2020-05-12 | 中国科学院半导体研究所 | Simplex Deep Neural Network Model and Construction Method Based on Homologous Continuity |
| CN111243021A (en) * | 2020-01-06 | 2020-06-05 | 武汉理工大学 | Vehicle-mounted visual positioning method and system based on multiple combined cameras and storage medium |
| CN111241986A (en) * | 2020-01-08 | 2020-06-05 | 电子科技大学 | A closed-loop detection method for visual SLAM based on end-to-end relational network |
| CN111241986B (en) * | 2020-01-08 | 2021-03-30 | 电子科技大学 | A closed-loop detection method for visual SLAM based on end-to-end relational network |
| CN111753789A (en) * | 2020-07-01 | 2020-10-09 | 重庆邮电大学 | Robot Vision SLAM Closed-loop Detection Method Based on Stacked Combined Autoencoders |
| CN113066152A (en) * | 2021-03-18 | 2021-07-02 | 内蒙古工业大学 | AGV map construction method and system |
| CN113066152B (en) * | 2021-03-18 | 2022-05-27 | 内蒙古工业大学 | AGV map construction method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109341703B (en) | 2022-07-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109341703A (en) | A full-cycle visual SLAM algorithm using CNNs feature detection | |
| CN111311666B (en) | Monocular vision odometer method integrating edge features and deep learning | |
| CN109800689B (en) | Target tracking method based on space-time feature fusion learning | |
| CN109816689A (en) | A moving target tracking method based on adaptive fusion of multi-layer convolutional features | |
| Zhong et al. | DynaTM-SLAM: Fast filtering of dynamic feature points and object-based localization in dynamic indoor environments | |
| CN112861808B (en) | Dynamic gesture recognition method, device, computer equipment and readable storage medium | |
| CN111310659A (en) | Human body action recognition method based on enhanced graph convolution neural network | |
| CN104200237A (en) | High speed automatic multi-target tracking method based on coring relevant filtering | |
| CN111161315A (en) | A multi-target tracking method and system based on graph neural network | |
| CN106909938B (en) | Perspective-independent behavior recognition method based on deep learning network | |
| Fei et al. | Flow-pose Net: An effective two-stream network for fall detection | |
| CN112699954B (en) | A closed loop detection method based on deep learning and bag-of-words model | |
| CN106023211A (en) | Robot image positioning method and system base on deep learning | |
| CN103761747B (en) | Target tracking method based on weighted distribution field | |
| CN113989920A (en) | Athlete behavior quality assessment method based on deep learning | |
| Wei et al. | GMSK-SLAM: A new RGB-D SLAM method with dynamic areas detection towards dynamic environments | |
| Yue et al. | SCFNet: Semantic correction and focus network for remote sensing image object detection | |
| Dong et al. | Weighted triplet loss based on deep neural networks for loop closure detection in VSLAM | |
| CN112432653B (en) | Monocular visual inertial odometry method based on point and line features | |
| Peng et al. | Visual SLAMBased on Object Detection Network: A Review. | |
| CN109766790A (en) | A Pedestrian Detection Method Based on Adaptive Feature Channel | |
| CN115880332B (en) | Target tracking method for low-altitude aircraft visual angle | |
| CN114140524B (en) | A closed-loop detection system and method based on multi-scale feature fusion | |
| Feng et al. | Effective venue image retrieval using robust feature extraction and model constrained matching for mobile robot localization | |
| Wu et al. | A novel gesture recognition model under sports scenarios based on kalman filtering and yolov5 algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |