Domestic Traffic Sign Recognition And Embedded Implementation Based On Yolov3

In recent years,with the breakthrough in the field of artificial intelligence,the introduction of auxiliary driving system has changed the previous driving mode.By acquiring real-time road information,the system can prompt the driver to make accurate operation in time,so as to prevent traffic accidents caused by fatigue driving.Traffic sign is an important road information,how to identify traffic sign quickly and accurately is very important.However,the traditional target detection algorithm faces some disadvantages in the real-world test,such as: easy to be limited by light,angle,obstacle occlusion,driving speed and other factors,and it is difficult to achieve multi-target detection,easy to miss detection,slow recognition.Aiming at the need of real-time and accurate identification of traffic signs,this paper takes targeted research and development,the main work is as follows:(1)In view of the small number of research samples of traffic signs in China,cctsdb data set with relatively complete data set is selected,and cctsdb is preprocessed and expanded.There are three types of annotation data: indication sign,prohibition sign and warning sign.Common traffic signs have been included,and the complexity of the environment selected in the picture is in line with the actual situation.(2)Firstly,we use the deep learning network yolov3 to train cctsdb.In the test,the accuracy of image recognition is over 98%,and the recognition time is about 0.03 s.It can be seen that yolov3 can be used for real-time detection and recognition.However,in video recognition,it is easy to miss detection or even wrong detection for small distant targets.In order to solve this problem,this paper makes a clustering analysis of the anchor value under the Yolo layer,combines the coordinate parameters of the ground truth of the experimental data set,and modifies the network parameters.The test results show that the performance of the optimized yolov3 is improved for the spatial location of small targets,and the recognition effect is improved.(3)In order to meet the needs of practical application and facilitate the transplantation of embedded system,this paper deals with the lightweight of the yolov3 model,taking the yolov3 tiny framework as the template,reducing the number of network layers to reduce the computational complexity.In the selection of development board,this paper uses the embedded AI development kit Jetson nano of NVIDIA.Its advantage is that it supports NVIDIA jetpack and can run multiple neural networks in parallel for image classification,target detection,segmentation and voice processing applications.In addition,it carries the CSI camera of raspberry pie.The test results show that the recognition time for imagesamples is about 0.1s,and the number of real-time recognition frames can reach 15 fps,which basically meets the conditions of real-time detection.