Design Of Road Vehicle Detection System Based On Deep Learning

With the rapid development of society and economy,the car ownership rate of residents has greatly increased,and transportation has become more convenient.At the same time,problems such as traffic congestion,frequent traffic accidents,and increased urban traffic pressure have gradually become prominent,which have become common problems faced by all countries in the world.For the traffic management department,accurately grasping the real-time traffic flow and realizing the real-time positioning of road vehicles will be conducive to timely traffic guidance and reduce the probability of traffic accidents.The road vehicle detection technology is also the basis for obtaining information such as the number and location of vehicles for tasks such as traffic flow statistics.Therefore,the research on vehicle detection technology has important practical significance and practical application value.Vehicle detection uses target detection technology.This paper introduces the traditional target detection algorithm and the target detection algorithm based on deep learning.Judging from the current development trend,the development of target detection algorithms based on deep learning is faster than that of traditional algorithms.In terms of hardware,compared with CPU(Central Processing Unit),GPU(Graphic Processing Unit),and ASIC(Application Specific Integrated Circuit),FPGA is more suitable for terminal deployment of deep learning networks.For example,Xilinx and Intel have the deep learning applications of FPGA products are supported by complete software and hardware development tools.Therefore,in order to meet the real-time,accuracy and convenience requirements of road vehicle detection scenarios,the paper adopts the single-stage target detection framework SSD based on deep learning as the network model,and selects the Xilinx Zynq Ultra Scale+MPSo C ZCU102 development board as the hardware deployment platform.A new road vehicle detection system based on FPGA is constructed in a semi-customized way based on the unified development platform of Vitis.On the technical route,this paper adopts the semi-custom development method based on the Vitis unified development platform.The core module of the convolutional neural network is implemented by the deeply optimized and customized Xilinx official DPU IP core.For the algorithm preprocessing and matrix postprocessing and decoding operations that still have a certain amount of calculation,considering that the use of the FPGA IP core can eliminate the speed bottleneck well,Therefore,the combination of DPU IP core and FPGA IP core is used to design the system hardware platform.Specifically,the DPU IP core and the FPGA acceleration IP core use the V++ instruction to compile and generate the overall hardware platform of the system through the Xilinx Vitis platform,and then the V++ instruction will link the overall hardware platform of the system,the official Xilinx embedded basic platform and the file system in the general ext4 format to generate the operating system image.In terms of hardware calls,this paper generates a DPU network inference function library based on Xilinx Vitis AI technology and designs a high-level comprehensive hardware scheduling function library based on Open CL technology,which are compatible with multi-threaded host applications program based on Python design in the form of dynamic link libraries.The " Vehicles traffic recordings dataset on public road " used for model training and testing comes from Mendeley Data’s official website.The author of the paper uses the dataset labeling tool Label Img to label the dataset to obtain labels in VOC format,and then convert the dataset images and corresponding labels into TFrecord format and send them to the algorithm network for training.The network model obtained by training is used to generate DPU dynamic link library and executable file,and we use SSH to connect the host and development board remotely to realize the overall debugging of the system.The test results show that the system can accurately detect and identify vehicles on the ZCU102 development board.It takes about 55.33 ms to process a 1280×720 picture by using the test method of reading pictures from SD card.The average computing power performance value of each network layer of the DPU is about 92.8GOP/S,the computing power of individual layers can reach a peak value of about 819.9GOP/S;and in the test mode of capturing images through a camera with a resolution of 640×480,the real-time processing speed of the system is about 17 frames per second,and the power consumption is 30.8W.The above test results show that the new road vehicle detection system designed in this paper meets the requirements of real-time detection of vehicles in highway scenes.The main work content and characteristics of the paper are: 1.A deep learning framework based on FPGA is constructed in a semi-custom development way,which provides a reference for the circuit deployment of deep learning;2.A road vehicle detection scheme based on deep learning is realized,which meets the requirements of real-time detection;3.A road vehicle dataset is marked,which provides support for other related work in this field.