Traffic Target Recognition Based On Convolution Neural Network And Zynq

In recent years,with the popularization of automobile in life,the research of Intelligent driver assistance technology has been paid more and more attention by all circles of society,and the ability of road target detection is the key link of practical intelligent auxiliary driving technology.It has become a hot research topic to use neural network to detect road targets in vehicle embedded.However,at present,efficiency of the mainstream ARM with neural network to identify the road target detection is not high.This paper focuses on the way of optimizing neural network on Zynq to improve the efficiency of road target detection.The main research contents include:(1)Road target detection: Research on the use of Tiny-YOLO model for the road target detection.In the course of the research,the training dataset is calibrated,and the number of fixed points with different digits is used instead of floating-point numbers,and the accuracy,speed and influence of the training are analyzed.(2)The Zynq realization of convolution neural network: This system uses HLS to design the forward propagation module,and the IP core of convolution neural network based on FPGA is realized,and then the fixed point optimization and pipeline optimization of convolution operation are carried out.(3)The realization of embedded system: Using the method of collaborative design of software and hardware,the Pynq project is transplanted to Zedboard.The image transmission module and the hardware convolution neural network IP Core are constructed in the programmable logic part of the Zynq chip.Finally,the three parts of image acquisition,road target detection and image display are integrated into an embedded system.This system uses the parallelization embedded Zynq to realize the quantization neural network,and completes the acquisition and detection function of the road target.Cars,pedestrians,traffic signs and non-motorized vehicles can be detected in system tests.Compared with the neural network which uses ARM alone with non-fixed point,it improves the efficiency of road target detection and provides a new realization scheme for intelligent auxiliary driving system.