Design Of Vehicle Target Detection System Based On FPGA And CN

With the rapid development of China’s economy,national car ownership is growing,bringing serious challenges to urban traffic.To alleviate traffic congestion and reduce the incidence of traffic accidents,Intelligent Transportation Systems(ITS)in China are developing rapidly.Vehicle detection is particularly important for real-time feedback of traffic information.In recent years,the performance of traditional methods for vehicle target detection has been mediocre,while deep learning-based vehicle target detection methods have gradually become mainstream algorithms.FPGAs,with their high parallelism and pipelined processing,not only have advantages in processing speed and power consumption,but also have flexible and convenient reconfigurable designs.Therefore,this paper is based on FPGA for vehicle target detection system design.This paper adopts CPU+FPGA heterogeneous computing platform to design the vehicle target detection system,and the CPU side is responsible for completing the design of convolutional neural network Mobile Net-based vehicle target detection algorithm,and transmitting the trained network parameters to FPGA through fiber optic module;the FPGA side uses logic circuit to build the network model structure and perform real-time vehicle target detection on the video images captured by camera OV5640.The FPGA side uses the logic circuit to build the network model structure,real-time vehicle target detection on the video images captured by camera OV5640,and the video images and vehicle target detection results are displayed on the display in real time through HDMI.The main research in this paper includes:(1)A top-down modular design of the system,pipelining and parallel processing of modules such as data communication,image acquisition and display control.(2)A universal convolutional computation engine is designed for the computation of all convolutional layers in Mobile Net,avoiding the separate design of different types of convolutional computation structures,and at the same time,multi-channel parallelism and resource cyclic reuse design for different types of convolutional layers,improving computation speed while reducing resource consumption.(3)According to the processing characteristics of FPGA,the lightweight convolutional neural network Mobile Net is selected as the basis of vehicle detection algorithm,and the width factor α and resolution factor β are reasonably set to compress and slim down the network model,and the network parameters are designed in a fixed-point manner.(4)The convolutional layer and batch normalization are designed for layer fusion to simplify the deployment of the network model in the FPGA and improve the computational efficiency.Finally,the system hardware design is completed based on Xilinx’s XC7K325 FPGA chip with Verilog hardware description language.For the hardware deployment of the optimized algorithm,the total system power consumption is 2.89 W at 50 MHz clock frequency,and the power consumption of the vehicle detection algorithm part is 0.688 W.The detection frame rate can reach 52.69 fps.Compared with the related literature in recent years,this design consumes less hardware resources and has certain advantages in terms of processing speed and power consumption.