Design And Implementation Of Vehicle-mounted Target Detection System Based On ZYNQ

With growing economic,increasing numbers of vehicles bring us convenience,but at the same time it brings us social problems which are much more outstanding these days such as environmental pollution,traffic jams and traffic accidents.Scholars across the world devotes to the study of smart traffic system in order to resolve above problems.The vehicle-based object detection system,as an important part of automatic driving system,can reduce traffic accidents effectively by monitoring target's type and position in its view to assist driver and vehicle-based control system make decision when driving.Recent years,a number of image detection algorithm innovations based on convolutional neural network structure have sprung up with the maturity of computer technology.Compared with the shallow machine learning algorithm,these new algorithms have greatly improved the detection accuracy.However,because of its complex structure and large amount of computation,it is difficult to meet the requirements of detection speed,detection accuracy,hardware power consumption and hardware cost in embedded real-time applications.In this paper,my research object is the convolutional neural network object detection which is transplanted to Xilinx ZYNQ(ARM + FPGA)scalable processing platform with different architecture by the method of combination of hardware and software.Firstly,to solve the problem of structure complexity,I improved the software and hardware division of convolutional neural network calculation based on the structural characteristics of YOLO convolutional neural network.This study adopted ARM Cortex-A9 processing module(PS)to carry non maximum inhibition algorithm,control algorithm flow and classify targets.Hardware acceleration is carried out in the programmable logic FPGA module(PL)by parallel computing in pooling layer and the convolution layer together.Secondly,to solve the problem of calculation delay and high resource utilization rate in the parallel computation of programmable logic,the convolution layer calculation is accelerated through array partitioning,loop unrolling,and loop pipelining.It divided data into 4 parts to transmit data at the same time by 4 AXI-HP ports in ZYNQ.It reduced reading and writing latency between FPGA and DDR memory.The experiment shows that the detection results are accurate under various complex road conditions and light conditions with good robustness in function.In terms of detection speed,the hardware acceleration method improves the operation throughput and efficiency when the core device power consumption is only 3W,and achieves speed of 10 frames per second.Compared with arm cortex-a9,the throughput is increased by 111 times and the operation efficiency is 57.5 times.In terms of detection accuracy,the accuracy and coverage of YOLOv2 algorithm in different detection scenarios are more than 80%.In a word,indexes of the system meet requirements as designed.Based on the ZYNQ hardware platform and convolutional neural network target detection algorithm,a low-power vehicle target detection system with low cost,realtime and high detection accuracy is designed and implemented.Not only can the research results be applied to the vehicle driver's dangerous target detection and alarm system,but also vehicle driving assistance system and driverless system in the future.