An FPGA Implementation Of Real-time Object Detection Algorithm Based On The YOLOv3 Model

With the increasing number of cars in the society and the increasing incidence of traffic accidents,the demand for vehicle visual assistant driving is increasing.Object detection is the core component of vehicle vision assisted driving system.In recent years,deep learning has developed rapidly,and the recognition accuracy and speed of object detection algorithm have been greatly improved.However,due to its complex structure,large amount of calculation and high requirements for hardware performance,the object detection algorithm based on deep learning is often difficult to achieve in engineering applications.This paper compares and analyzes the advantages and disadvantages of the current object detection algorithm based on deep learning.Compared with other mainstream object detection algorithms,YOLO series algorithm has the advantages of faster detection speed and better real-time performance.In the aspect of hardware implementation,compared with CPU,GPU and ASIC embedded hardware,FPGA has the advantages of low power consumption,programmability,parallelism and low cost.Aiming at the practical engineering application of the object detection algorithm in intelligent driving,this paper selects the appropriate embedded low-power hardware and deploys the best object detection algorithm.At the same time,we uses DP-8020 development board and DNNDK deep compression tool developed by Shenjian technology company to realize the real-time object detection of Mobile Net-YOLOv3 object detection algorithm on FPGA.The main work is summarized as follows:(1)In order to make the object detection algorithm more practical,this paper first uses the A3 D infrared thermal imager of Infiray Company to collect video data,and self-made 50000 infrared image data sets(resolution of 384×288 pixels)for the training and testing of algorithm model.(2)Starting from the feasibility of the practical application of the object detection algorithm,taking the algorithm model of YOLOv3,Tiny-YOLOv3 and Mobile Net-YOLOv3 as the research object,this paper analyzes the important hyperparameters that affect the accuracy of the model,and focuses on adjustment and optimization of three hyperparameters of Learning rate,Epoch and Batch＿size to make the algorithm model and the data set more matching and improve the training effect.Finally,Mobilenet-YOLOv3 algorithm is selected as the best object detection algorithm and transplanted to FPGA platform through compare and analysis the mean average precision(m AP)and frame rate of three algorithms.(3)Using DNNDK deep compression tool to compress and optimize the Mobile Net-YOLOv3 algorithm model,then using Vivado software to build the FPGA hardware module,and using Peta Linux system to build DP-8020 platform embedded Linux system.Next,using Linux SDK software to create algorithm test executable program.Finally,migrates the system and files to FPGA to complete the transplantation and function realization of object detection algorithm.The experimental results show that the mean average precision(m AP)made by real-time detection of Mobile Net-YOLOv3 algorithm based on FPGA platform can reach 82.50%,and the frame rate can reach 38 fps,which can meet the the real-time object detection requirements of engineering applications,and has great advantages and value in engineering applications.