Pedestrian Crossing Red Light Detection Based On Deep Convolutional Neural Network

The government has been urging people to travel in a civilized manner to reduce unnecessary injuries and death rates,but some people and cars still run red lights when crossing crosswalks.Pedestrian and vehicle have become the focus in the field of target detection,and target detection is widely used in traffic,intelligent monitoring and vehicle-assisted driving.Among target detection algorithms,YOLO V5 algorithm is one of the most advanced algorithms in pedestrian detection.Therefore,in view of this problem,it is particularly necessary to study a set of red light detection system for pedestrians and vehicles.This paper uses deep learning YOLO V5 algorithm combined with Deep Sort to detect pedestrians running red lights.The specific work is summarized as follows:(1)First of all,the convolutional neural network theory and YOLO series algorithms involved in target detection are described in detail in this paper,which lays a foundation for selecting YOLO V5 network structure as the development of pedestrian red light running detection in this project.(2)Secondly,this paper roughly describes the network framework CNN for basic deep learning,and describes the network structures Alex Net,VGG and Rest Net that evolved from CNNs with faster speed and better extraction features,and describes the network optimization training method.The general structure of YOLO and the detailed structure of YOLO V5 are described at the same time,and the advantages of each functional module are described.The COCO dataset was used to test the mainstream object detection algorithms,and first the YOLO V3 and the traditional algorithm were compared to obtain that the YOLO V3 algorithm performed better,and then the YOLO V3,V4,V5 were compared,of which YOLO V5 was better expressive.(3)Then,improve YOLO V5 network,add Retinex fogging image enhancement algorithm in the head,and transform the adaptive frame of Backbone terminal into direct secondary sampling extraction feature,light code,improve the image blur caused by weather reasons such as night and rain,change the number of end output parameters,From the original model 80 species to 5 species.The accuracy rate is 94%,recall rate is 90%,m AP is 80.27%,FPS is 19,which is better than the traditional model.The improved YOLO V5 algorithm has good effect.Deep Sort algorithm was added to track pedestrians to prevent the same pedestrian from being identified as another person for a long time.(4)Finally,a complete project is made,the principle of each functional module is described,the project of detecting pedestrians running red lights is completed,the project is summarized and the future is looked forward to.This project is committed to making a code lightweight,in different weather to make the pedestrian recognition rate higher project module,and can be applied to the actual project.