Research On Multiple Object Tracking Under Complex Scene

Multi-object tracking is an important but challenging task in computer vision.It is widely used in automated surveillance,robotics,radar-based tracking of aircraft,human-computer interaction,sports video analysis and video detection and recognition.With the improvement of object detection technology,the tracking-by-detection methods become popular.The tracking-by-detection multiple object tracking methods detect the objects in video and find the locations of objects first,then the data association algorithm is used to associate these detection responses into long trajectories.There are many research directions in multi-object tracking area,such as 3D multi-object tracking,multi-camera multi-object tracking and so on.The research works in this paper are focused on single camera multi-object tracking on image plane,and the main contents are listed as follows:First,aiming at the problem that the discrete-continuous energy minimization(DCEM)method cannot effectively deal with the trajectory fragmentation and identity switch in complex scene,an improved DCEM multi-target tracking method is put forward.In order to solve identity switch,this method extracts the multi-feature fusion appearance vector of the tracked target,and uses the Euclidean distance between appearance vectors from different targets to design the constraint function for trajectories.Then this method designs post-processing process for trajectory fragmentation problem by merging the short tracklets which have high similarity of motion and appearance in adjacent spatial-temporal neighborhood.Experimental results indicate that the problems of trajectory fragmentation and identity switch can be greatly improved.Secondly,aiming at the problem of the complex network structure and partial occlusion in the video based on network flow tracking method,a tracking-by-detection multi-target algorithm is proposed based on network flows using hierarchical data association for multi-target tracking in complex environment.Firstly,a dual-threshold strategy is adopted to perform low-level association and generate reliable tracklets between adjacent frames according to detection responses.Then,based on the directed acyclic graph formed by tracklets,the minimum cost flow algorithm is used to obtain the long trajectories.In order to deal with occlusion problem,the proposed method adopts part-based appearance models during the process of association.Experiments show that the proposed method can handle occlusion and trajectory fragment effectively during the tracking process in complex environment.Finally,aiming at the problem that the motion of target is always non-linear in real-world surveillance video especially when there is occlusion,a affinity measurement scheme with a nonlinear motion model is proposed for multi-object tracking.In this method,each node in network represents a tracklet,and each edge represents the likelihood of neighboring tracklets belonging to the same trajectory as measured by the proposed affinity score.When the affinity score is measured,the object appearance and motion information are used to measurement.The appearance information is obtained by the robust multi-feature fusion and the motion information is obtained using the linear motion model and the nonlinear motion model.Besides,after global optimization,the proposed nonlinear motion model is used to fill the vacancies in trajectories.Experimental results demonstrate the proposed method can achieve continuous tracking trajectories under the case of motion direction change and complete occlusion.