Research On Vehicle And Pedestrian Detection And Tracking Algorithm Based On Improved YOLOv5

In recent years,artificial intelligence technology has developed rapidly and has gradually penetrated into the transportation industry.Using deep learning technology to detect and track vehicles and pedestrians in real time is an important part of autonomous driving technology.Pedestrian safety,so as to avoid the occurrence of traffic accidents.However,in real driving scenarios,the environment is highly complex,and the existing algorithms have slow detection speed,low precision and poor tracking effect.In view of this,this paper proposes an improved YOLOv5 algorithm to detect vehicles and pedestrians,and combines with two tracking algorithms to track pedestrians and vehicles respectively.Specifically,the following research is carried out:(1)This paper selects the public dataset SODA10 M and divides the dataset categories into five categories: Pedestrian,Cyclist,Car,Truck,and Tram,and uses flipping and rotating methods for data enhancement,which solves the problems of uneven sample distribution in the original dataset to a certain extent.(2)YOLOv5s and YOLOv3 were used to conduct experiments on the SODA10 M dataset respectively.In the same experimental environment,the final results showed that the detection accuracy of the YOLOv3 algorithm was better than that of YOLOv5 s,but in terms of model size and detection speed,the YOLOv5 s model was smaller and the detection speed was faster.(3)Under long-distance detection,vehicles and pedestrians occupy fewer pixels in the picture,making it difficult to distinguish them from the surrounding background,resulting in the inability of the neural network to extract effective information during the training process.In this regard,this paper proposes an improved model based on YOLOv5 s.First,an attention mechanism is added to the backbone feature extraction network of YOLOv5 s to eliminate interference with irrelevant information.In this paper,four attention mechanisms: SE,ECA,CBAM and CA are added to the backbone network.Experiments show that CBAM performs better;secondly,in the backbone feature extraction network and detection head,the small target feature layer is introduced to increase the receptive field of the network;finally,the original PANet feature fusion method is changed to Bi FPN feature fusion to further enhance feature fusion of different dimensions.The final improved model achieved 85.2% m AP on the experimental dataset,which was 2.7% higher than YOLOv5 s,and the FPS reached 42,which was 7s lower than YOLOv5 s.(4)The improved YOLOv5 s is used as the target detector of Deep SORT and Byte Track two tracking algorithms;The pedestrian re-identification data set Market-1501 and the vehicle re-identification data set Veri-776 are used to train the appearance feature extraction network in Deep SORT,training pedestrian and vehicle feature extraction models;the tracking effect of the two tracking algorithms in the pedestrian scene was tested in the three scenarios of the MOT16 pedestrian tracking data set;In terms of vehicle tracking,a surveillance video of traffic roads was intercepted and the video was split into video frames,a vehicle tracking test set was established using Darklabel annotation software.Experimental results show that compared with Byte Track,the improved YOLOv5 algorithm combined with Deep SORT can stably track pedestrians and vehicle targets on the road to a certain extent,and has certain practical value.