Research On Vehicle Target Detection And Tracking Technology Based On YOLOv3

Compared with fixed monitoring equipment,UAVs have the advantages of low price,convenient deployment,and flexible mobility,and have broad application prospects in traffic information collection.For this reason,in the context of the complex road conditions of ordinary urban roads,based on the YOLOv3 algorithm,the vehicle target and tracking technology from an aerial perspective is studied in this paper,and Optimization improvements made in reducing computing resources,speed up and enhance the precision.In terms of vehicle detection,in order to improve the application effect of YOLOv3 algorithm in aerial vehicle detection,YOLOv3-Aerial aerial vehicle target detection algorithm is proposed in this paper.A total of five improvement measures are proposed for the default YOLOv3 from the aspects of network structure and loss function.Firstly,the layers and filters of the Darknet basic network are appropriately pruned,and the detection speed is greatly improved with a limited reduction in accuracy.Secondly,an improved spatial pyramid pooling is introduced to pool and connect multi-scale local area features so that the network can learn the target features more comprehensively.Then,an improved FPN network design is proposed,which effectively improves the detection effect on small targets of non-motor vehicles.Furthermore,the focus loss function is used instead of the cross-entropy loss function to solve the problems of false detection and false alarm caused by the imbalance of the number of sample categories in the aerial vehicle data set.Finally,GIOU is used to replace the mean square error regression loss function,which alleviates the problems of detecting the position deviation and the inaccurate overlap of the bounding box.In terms of vehicle tracking,an improved SORT multi-target tracking algorithm is designed which fuses appearance features.Aiming at the situation that the default SORT multi-target tracking algorithm has frequent target ID switching and small target association errors between frames,a data association matrix that combines improved spatial color histogram appearance features and GIOU spatial location features is proposed instead of the default IOU similarity matrix.Thereby,the Hungarian algorithm can better match small targets between dense non-motor vehicle frames and improve tracking accuracy.In summary,the above improvement measures can increase the detection accuracy by3.51%,increase the detection speed by more than 3 times,increase the tracking accuracy of small targets by about 2%,and meet the needs of real-time detection and tracking of aerial video.In addition,It has certain theoretical significance and practical application value in related fields such as aerial vehicle detection and tracking.