Design And Optimization Of Object Detection Algorithm On FPGA Computing Platform For Autonomous Vehicle

Object detection is one of the crucial technologies of self-driving environment sensing.However,the complexity and numerous parameters of of the algorithm challenge the computing resources and the power consumption of the on-board computing platform.Hence,how to design the lightweight neural network that meets the requirements of hardware characteristics is a difficult problem facing the current autonomous driving perception system.Therefore,considering the algorithm and computing-platform characteristics comprehensively,one kind of the implementation and optimization of object detection based on FPGA computing platform for autonomous vehicle was proposed.Under the limit of low FPGA power consumption,considering the dilemma between the current algorithm and the computing-platform characteristics,a kind of joint learning algorithm of object detection and computing-platform was presented,and a lightweight backbone network was designed.Based on the backbone network and the complexity of the driving images,a dynamic network architecture of object detection algorithm based on image preclassification is proposed to achieve high computational efficiency.In this paper,Xilinx ZCU102 is used to verify the above methods and theories,and the computing-platform acceleration method is used to implement the efficient selfdriving sensing system.The main work contents are as follows:(1)In view of the separation dilemma between algorithm design and computingplatform design,a joint learning algorithm of object detection and computing-platform was proposed to design a hardware-friendly network,in consideration of the hardware characteristics.The neural network architecture search algorithm was used in the design process,and the algorithm was improved to solve the problem of tremendous search space and longstanding search.Through the statistical analysis of the architecture parameters of the existing efficient CNN,the rules of the five parameters,including the amount of parameters,the number of layers,the resolution,the settings of network blocks and the convolution kernel,and the classification accuracy are obtained.Combined with the analysis of the characteristics of FPGA hardware,the search space was effectively reduced.In terms of search methods,based on the joint learning algorithm of object detection and computing-platform,a search algorithm combining random search,group supervised search and FPGA feature prediction is adopted to obtain an efficient lightweight backbone network SHNet,and the classification accuracy of Top-1 in Image Net is 77.2%.(2)In order to further utilize computing resources efficiently,a dynamic network architecture object detection algorithm based on image pre-classification is presented through image information with different complexity,to improve the real-time performance and accuracy of object detection.In this paper,a kind of image complexity classification method is presented based on object occlusion,image truncation,size of bounding box,total number of targets and the omission rate of representative networks in a single frame image,and a complexity pre-classification model for image object recognition is obtained.Based on the lightweight backbone neural network mentioned above,the dynamic detection architecture DSHnet was designed,which reached 82.14%m AP@0.5 on VOC2007 test dataset,and the depth-wise convolution was adopted in the detection neck to reduce the computational load and realize the high efficiency of operation.(3)In order to carry out experimental verification of the optimized lightweight object detection,this paper uses Xilinx compiler tool to implement model quantization processing and compilation of the DSHnet model on ZCU102.There were 100 images were sampled randomly from the VOC2007 test dataset for the speed test,and the average frame rate is 43 frames/s.The efficient self-driving sensing system was realized,and the validity of the above method and results is verified in the paper.