Research On Multi-target Multi-category Tracking Algorithm Based On DeepSort In Complex Traffic Scene

With the rapid development of society,people have increasingly strong diversified demands for traffic travel,which leads to the increasing complexity of traffic.In complex traffic environment,traffic congestion and traffic safety problems are easy to occur,and it is a very necessary and difficult task to accurately obtain the traffic information of the scene.In the field of computer vision,object tracking is a hot research direction,to the field of traffic could be obtained by visual sensor real-time traffic conditions,make traffic for timely decision to provide technical theoretical support to avoid the occurrence of the traffic problem.This paper mainly selects the detection based object tracking strategy,and uses the representative simple online and real-time depth association measurement algorithm(Deep Sort),based on which,a multi-object and category tracking algorithm for complex traffic scenes is proposed.In the object detection part,Vis Drone data set is used,which contains ten common object categories in urban traffic scenes and is filmed by UAV,providing data support for the study of multi-object and multi-category tracking algorithm in this scene.The data set is applied to the experimental part of object detection and object tracking algorithm.object detection is the first step to obtain object information in the scene,and its performance directly affects the tracking algorithm.In this paper,the Faster-RCNN,YOLO v3 and YOLO v5 detection algorithms are compared,among which YOLO v5 has four network models ranging from small to large: YOLO v5 s,YOLO v5 m,YOLO v5 l and YOLO v5 x were trained and tested on Vis Drone data set to comprehensively evaluate the detection accuracy and speed of various algorithms.YOLO v5 m model was selected as the detector of tracking algorithm.In the part of object tracking,the current tracking algorithm cannot accurately track the object category information in complex traffic scenes,which is meaningless.Aiming at the lack of discrimination of object category information in the current multi-object tracking strategy,a multi-object multi-category tracking algorithm is proposed.The original apparent model of Deep Sort algorithm was trained based on pedestrian data set.For features of other categories with poor feature extraction ability,data set was constructed to retrain the apparent model and improve the accuracy of the algorithm for tracking object location information.The classification information is integrated into Kalman filter,and the state variables and observation variables are defined again.Then,based on the category co-occurrence matrix obtained by Deep Sort tracking,Jaccard index was used to construct the category similarity network.Combined with the network,the confidence change when the error detection occurred in the detection category was statistically analyzed,and the dynamic threshold function was obtained by curve fitting.The category matching module was constructed to complete the category matching and correction tasks.Improve the accuracy of the algorithm to object category information tracking.After experimental verification,compared with YOLO v5m+Deep Sort algorithm,the multi-object and category tracking algorithm established in this paper improves the Multi object tracking accuracy(MOTA)by 5.5%,the Multi object tracking category accuracy(MOTCA)by 8.3%,and the object center accuracy under the threshold of 25 by 5% in complex traffic scenes.The overlap rate of threshold 0.5 was increased by 9%.The results show that the multi-object and multi-category tracking algorithm can more accurately and perfectly complete the object tracking task in complex traffic scenes,which is of great significance to the development of intelligent transportation system.