Research On Real-time Tracking Method Of People And Vehicles Based On Bilateral Association

Multi-object tracking is an important branch of computer vision,which is mainly used in behavior recognition,event analysis,video surveillance and other fields.At present,most of the research focuses on the accuracy of tracking,but real-time performance is also urgently needed and lack of research.Tracking pedestrians and vehicles in real time is very important to realize smart cities and traffic safety.Therefore,this paper focuses on the multi-object tracking task in complex scenes,and studies the real-time tracking method of pedestrians and vehicles.The main research contents of this paper are as follows:(1)Focusing on the general dependence of detection-based tracking algorithms on detectors,a Hierarchical Agglomerative Clustering(P-HAC)algorithm based on density estimation of Parzen is proposed to generate tracklet of basic points of tracking.The algorithm avoids the excessive calculation caused by traditional deep network or global optimization algorithm in the detection association stage,and at the same time alleviates the influence of detector fault on tracking performance,and improves the internal accuracy of tracklet.Therefore,P-HAC algorithm is used to form tracklet by association detection,which is used to solve the general dependence on detectors in real-time tracking.(2)Aiming at the problem that tracklet features have poor description ability to features,a robust Dual Appearance Feature(DAF)construction method is proposed to describe the current state and the changing trend of tracklet.This method endows tracklets with time sequence characteristics,reduces the drift of tracklet features caused by cross-frame,and enhances the robustness of tracklet features,and improving the accuracy of association between tracklets.(3)Aiming at the balance between accuracy and efficiency of real-time multi-object tracking algorithm,a Bilateral Association Tracker(BAT)based on tracklet is proposed.BAT has achieved excellent mixed tracking performance of people and vehicles under multiple benchmark datasets and different detection resolutions.The framework tracks the image sequence batch by batch,and obtains the detected motion and appearance information by using Kalman filter and ResNet50-ibn recognition network respectively,and then generates the tracklet as the tracked basic point by using P-HAC.In tracklet association,dual appearance feature DAF is used to enhance the ability of tracklet representation,and finally Hungarian algorithm is used to associate tracklets.On the popular benchmarks such as MOT2017,Visdrone and KITTI,the tracking accuracy of BAT is better than that of Deep-Simple Online and Real-time Tracking(DeepSORT),and there is no obvious decrease in efficiency.Compared with state of the art trackers,BAT has very competitive tracking accuracy and outstanding advantages in calculation cost.