CN110197497A - A kind of field biology tracing system and method based on deep learning - Google Patents

Abstract

The invention discloses a system and method for tracking wild creatures based on deep learning, belonging to the technical field of intelligent computing. The field biological tracking system based on deep learning of the present invention includes a video acquisition module, a video preprocessing module, a solar charging module, a target detection module, a wireless data transmission module and several GPS modules, and uses a data set to train a mobilenetv2-ssd model using a caffe framework , convert the trained mobilenetv2-ssd model into a model suitable for NCNN, deploy the frame-making and target detection modules of the target detection module to the PS side of the FPGA with the ARM core, and deploy the computationally intensive task convolution to the PL end. The deep learning-based wild animal tracking system of the invention can adapt to long-term field work and data transmission work, and has high stability and accuracy, and has good promotion and application value.

Description

A system and method for tracking wild animals based on deep learning

technical field

The present invention relates to the technical field of intelligent computing, and specifically provides a deep learning-based wild creature tracking system and method.

Background technique

FPGA (Field Programmable Gate Array), that is, Field Programmable Gate Array, is a product of further development on the basis of programmable devices such as PAL, GAL, and CPLD. It emerged as a semi-custom circuit in the field of application-specific integrated circuits (ASIC), which not only solves the shortcomings of custom circuits, but also overcomes the shortcomings of the limited number of original programmable device gates.

Heterogeneous computing with a reconfigurable architecture of CPU+FPGA has many advantages, such as: higher performance, greater flexibility, lower power consumption, inherent fault tolerance, and greatly shortened product development cycles. Using FPGA to replace GPU as the accelerator for future high-performance computing should be the main theme of the development of FPGA heterogeneous intelligent computing at this stage.

Deep learning refers to a collection of algorithms that use various machine learning algorithms to solve various problems such as images and texts on a multi-layer neural network. Compared with traditional algorithms, the accuracy of image recognition algorithms using deep learning is greatly improved.

Contents of the invention

The technical task of the present invention is to address the above existing problems and provide a deep learning-based wild animal tracking system that can adapt to long-term field work and data transmission work, and has high stability and accuracy.

The further technical task of the present invention is to provide a method for tracking wild animals based on deep learning.

To achieve the above object, the present invention provides the following technical solutions:

A wild animal tracking system based on deep learning, the system includes a video acquisition module, a video preprocessing module, a solar charging module, a target detection module, a wireless data transmission module and several GPS modules, using the data set to train mobilenetv2-ssd using the caffe framework model, convert the trained mobilenetv2-ssd model into a model suitable for NCNN according to the requirements of the embedded framework NCNN, deploy the frame-making and target detection modules of the target detection module to the PS end of the FPGA with an ARM core, and calculate Intensive task convolution is deployed to the PL side, while the video acquisition module, video preprocessing module, solar charging module, wireless data transmission module, and several GPS modules are deployed to the PS side.

Preferably, the wild creature tracking system based on deep learning also includes a system error recovery module and a system alarm module, and the system error recovery module and the system alarm module are deployed on the PS side.

Preferably, the video collection module and the video preprocessing module perform front-end preprocessing on the images collected by the camera to complete image enhancement and image compression.

Preferably, the wireless transmission module transmits the results of the target acquisition module to the cloud, and when no target is detected, the clock of the wireless transmission module is turned off.

Preferably, the solar charging module provides additional power supply for the system, and the solar charging module is always on.

A method for tracking wild animals based on deep learning. For specific wild animals, collect and label wild animal data, obtain the dataset with the tagged wild animals, use the caffe framework to train the mobilenetv2-ssd model, and use the mobilenetv2-ssd model to convert The tool is converted to the fixed-point reasoning framework NCNN, and then the frame-making and target detection modules are compiled and deployed to the PS side of the FPGA with multi-core ARM, and the calculation-intensive task convolution is deployed to the PL side, and the frequency acquisition module , video preprocessing module, solar charging module, wireless data transmission module, and GPS module are also deployed on the PS side.

The method for tracking wild creatures based on deep learning is realized by the system for tracking wild creatures based on deep learning described in the present invention. A wild animal tracking system based on deep learning, including a video acquisition module, a video preprocessing module, a solar charging module, a target detection module, a wireless data transmission module, several GPS modules, a system error recovery module and a system alarm module, using data sets using caffe The framework trains the mobilenetv2-ssd model, converts the trained mobilenetv2-ssd model into a model suitable for NCNN according to the requirements of the embedded framework NCNN, deploys the target detection module to the PS side of the FPGA with an ARM core, and converts the computationally intensive The task convolution is deployed to the PL side, and the video acquisition module, video preprocessing module, solar charging module, wireless data transmission module, several GPS modules, system error recovery module and system alarm module are all deployed to the PS side.

As a preference, a system error recovery module and a system alarm module are also used in the method, and the system error recovery module and the system alarm module are used to improve the stability of the system and are convenient to work in the field for a long time, and the system can be restarted through the cloud. Control, the system The error recovery module and system alarm module are deployed to the PS side.

Preferably, the video acquisition module and the video preprocessing module perform front-end preprocessing on the images collected by the camera to complete image enhancement and image compression.

Preferably, the wireless transmission module transmits the results of the target acquisition module to the cloud, and when no target is detected, the clock of the wireless transmission module is off; the solar charging module provides additional power supply for the system, prolonging the system. During working hours, the solar charging module is always on.

The deep learning-based wild animal tracking method first puts the trained model parameters into the system SD card. After the system is powered on, the video acquisition module and preprocessing module perform front-end preprocessing on the collected images to complete image enhancement and image compression. And a series of functions, at the same time compress the image to a size suitable for the mobilenetv2-ssd algorithm, the PS side of the system processor will read the model parameters from the SD card and send them to the PL side through DMA and data in parallel for algorithm acceleration to complete the calculation For intensive tasks, turn on the multi-core and multi-threading in ARM at the same time and use the neon accelerator to accelerate the frame-making task to complete the final target detection of the target wild animal. In order to reduce power consumption, only the data with the target wild animal will be transmitted through the wireless transmission module When transmitting to the cloud, when no target is detected, the clock of the wireless transmission module is turned off, and at the same time, in order to prolong the working time of the system, the solar charging module should always be turned on. The system alarm module and system error recovery module are used to improve the stability of the system to facilitate long-term field work, and at the same time, the system can be restarted and controlled through the cloud.

Compared with the prior art, the deep learning-based wild animal tracking system of the present invention has the following outstanding beneficial effects: the deep learning-based wild animal tracking system can adapt to long-term field work and data transmission work, and the stability The accuracy and accuracy are high. The system alarm module and system error recovery module are used to improve the stability of the system to facilitate long-term field work. At the same time, the system can be restarted and controlled through the cloud, which has good promotion and application value.

Description of drawings

Fig. 1 is a topological diagram of the deep learning-based wild animal tracking system of the present invention.

Detailed ways

The deep learning-based wild animal tracking system and method of the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments.

Example

As shown in Figure 1, the deep learning-based field biological tracking system of the present invention includes a video acquisition module, a video preprocessing module, a solar charging module, a target detection module, a wireless data transmission module, two GPS modules, and a system error recovery module and system alarm module.

The deep learning-based wild animal tracking system uses the data set to train the mobilenetv2-ssd model with the caffe framework, converts the trained mobilenetv2-ssd model into a model suitable for NCNN according to the requirements of the embedded framework NCNN, and converts the target detection module into a model The frame and object detection module is deployed to the PS side of the FPGA with ARM core, and the calculation-intensive task convolution is deployed to the PL side. At the same time, the video acquisition module, video preprocessing module, solar charging module, wireless data transmission module, two A GPS module, a system error recovery module and a system alarm module are deployed to the PS side.

The video acquisition module and the video preprocessing module perform front-end preprocessing on the images collected by the camera to complete image enhancement and image compression.

The wireless transmission module transmits the result of the target acquisition module to the cloud, and when no target is detected, the clock of the wireless transmission module is turned off.

The solar charging module provides additional power supply for the system, and the solar charging module is always on.

The system alarm module and system error recovery module are used to improve the stability of the system to facilitate long-term field work, and at the same time, the system can be restarted and controlled through the cloud.

The working process of the wild animal tracking system based on deep learning is as follows: First, put the trained model parameters into the system SD card, and after the system is powered on, the video acquisition module and preprocessing module perform front-end preprocessing on the collected images to complete A series of functions such as image enhancement and image compression, while compressing the image to a size suitable for the mobilenetv2-ssd algorithm, the PS end of the system processor will read the model parameters from the SD card and send them to the PL end in parallel through DMA and data. The algorithm accelerates the completion of computationally intensive tasks. At the same time, the multi-core and multi-threading in ARM is turned on and the neon accelerator is used to accelerate the frame-making task to complete the final target detection of the target wild animal. In order to reduce power consumption, only the data with the target wild animal It is transmitted to the cloud through the wireless transmission module. When no target is detected, the clock of the wireless transmission module is turned off. At the same time, in order to prolong the working time of the system, the solar charging module should always be turned on. The system alarm module and system error recovery module are used to improve the stability of the system to facilitate long-term field work, and at the same time, the system can be restarted and controlled through the cloud.

The wild animal tracking method based on deep learning of the present invention collects and labels wild animal data for specific wild animals, obtains the data set of the wild animal with labels, uses the caffe framework to train the mobilenetv2-ssd model, and uses the mobilenetv2-ssd model The conversion tool is converted to the fixed-point reasoning framework NCNN, and then the frame-making and target detection modules are compiled and deployed to the PS side of the FPGA with multi-core ARM, and the calculation-intensive task convolution is deployed to the PL side, and the frequency acquisition Modules, video preprocessing modules, solar charging modules, wireless data transmission modules, and GPS modules are also deployed on the PS side.

The method for tracking wild creatures based on deep learning is realized by the system for tracking wild creatures based on deep learning described in the present invention. A wild animal tracking system based on deep learning, including a video acquisition module, a video preprocessing module, a solar charging module, a target detection module, a wireless data transmission module, several GPS modules, a system error recovery module and a system alarm module, using data sets using caffe The framework trains the mobilenetv2-ssd model, converts the trained mobilenetv2-ssd model into a model suitable for NCNN according to the requirements of the embedded framework NCNN, and deploys the framing and target detection modules of the target detection module to the FPGA with the ARM core On the PS side, the calculation-intensive task convolution is deployed to the PL side. At the same time, the video acquisition module, video preprocessing module, solar charging module, wireless data transmission module, several GPS modules, system error recovery module and system alarm module are deployed to the PL side. PS side.

The system error recovery module and system alarm module are used to improve the stability of the system and facilitate long-term field work. At the same time, the system can be restarted and controlled through the cloud. The system error recovery module and system alarm module are deployed to the PS side.

The video acquisition module and the video preprocessing module perform front-end preprocessing on the images collected by the camera to complete image enhancement and image compression.

The wireless transmission module transmits the result of the target acquisition module to the cloud. When no target is detected, the clock of the wireless transmission module is off; The module is always on.

The solar charging module provides additional power supply for the system, and the solar charging module is always on.

The above-described embodiments are only preferred specific implementations of the present invention, and ordinary changes and replacements performed by those skilled in the art within the scope of the technical solution of the present invention should be included in the protection scope of the present invention.

Claims (9)

1. a kind of field biology tracing system based on deep learning, it is characterised in that: the system includes video acquisition module, view Frequency preprocessing module, solar recharging module, module of target detection, wireless data transfer module and several GPS modules, using number According to collection using caffe frame training mobilenetv2-ssd model, by trained mobilenetv2-ssd model according to embedding The requirement for entering formula frame NCNN is converted into being suitable for the model of NCNN, and module of target detection is played frame and module of target detection portion The realization of computation-intensive task convolution is deployed to the end PL, while video acquisition mould by the end PS for affixing one's name to the FPGA with ARM core Block, video pre-filtering module, solar recharging module, wireless data transfer module, several GPS modules are deployed to the end PS.

2. the field biology tracing system according to claim 1 based on deep learning, it is characterised in that: further include system Error recovery module and system alarm module, system mistake recovery module and system alarm module are deployed to the end PS.

3. the field biology tracing system according to claim 2 based on deep learning, it is characterised in that: the video is adopted Collect module and video pre-filtering module and camera acquired image is subjected to front end pretreatment, completes image enhancement and image pressure Contracting.

4. the field biology tracing system according to claim 3 based on deep learning, it is characterised in that: the wireless biography The result of target acquisition module is transmitted to cloud by defeated module, and when not detecting target, the clock of wireless transport module is Closed state.

5. the field biology tracing system according to claim 4 based on deep learning, it is characterised in that: the solar energy Charging module provides additional power supply for system, and solar recharging module is in the open state always.

6. a kind of field biology method for tracing based on deep learning, it is characterised in that: this method is any using claim 1-5 Field biology tracing system based on deep learning described in item claim is realized, for specific wild animal, is acquired and is marked Wild animal data are infused, the data set of the wild animal with label is obtained, utilize caffe frame training mobilenetv2- Ssd model is transformed on the Framework for Reasoning NCNN of fixed point using mobilenetv2-ssd model conversation tool, then will wherein Play frame and module of target detection compiling is deployed to the end PS of the FPGA with multicore ARM, computation-intensive task convolution is real Now be deployed to the end PL, with time-frequency acquisition module, video pre-filtering module, solar recharging module, wireless data transfer module, GPS module is also deployed to the end PS.

7. the field biology method for tracing according to claim 6 based on deep learning, it is characterised in that: in this method also Using system mistake recovery module and system alarm module, system mistake recovery module and system alarm module are used to improve system Stability convenient for for a long time in field work, while restart control, system mistake recovery to system by cloud Module and system alarm module are deployed to the end PS.

8. the field biology method for tracing according to claim 7 based on deep learning, it is characterised in that: video acquisition mould The progress front end pretreatment of camera acquired image is completed image enhancement and compression of images by block and video pre-filtering module.

9. the field biology method for tracing according to claim 8 based on deep learning, it is characterised in that:

The result of target acquisition module is transmitted to cloud by the wireless transport module, wireless to pass when not detecting target The clock of defeated module is in off state；The solar recharging module provides additional power supply for system, extends system Working time, solar recharging module are in the open state always.