Multiple Object Tracking With Motion Constraint

Object tracking is an important research field of computer vision,and multiple object tracking is one challenging research branch in object tracking.Multiple object tracking has a wide application in industry.With the rapid development,Multiple object tracking technology has achieved successes in many tasks,including autopilot,smart navigation,video surveillance,medical image processing and so on.In recent years,as deep learning methods have achieved breaking through in object detection,Tracking by detection has attracted lots of attention.The tracking by detection approach consists of two steps: object detection and data association.First,localize the objects in images with a pre-trained model.Then,associate the detections with a tracking algorithm.Finally,form trajectories of the targets.In this paper,we focus on the study of tracking by detection approach.In summary our contributions are:1.We have reviewed numerous multiple object tracking methods.Especially in related work section,we explained the theory and algorithm of the tracking by detection approaches in detail.2.In this paper,we proposed an online multiple object tracking algorithm based on structural motion constraints and Kalman filter.We adopt a two-step framework for online MOT,including tracking frame by frame and trajectory recovery.We leveraged two independent tracking models for each step,since each step has its special problem.3.We propose a motion structure to exploit the relative movements between objects in one same frame or adjacent frames,which could improve the discriminative ability of tracking algorithm.4.We propose a frame-by-frame tracking algorithm with the motion structure,which could supervise and constrain the data association process.5.Multiple object tracking algorithms suffer from mis-detections and false positives,which lead to trajectory fragments.In this paper,we propose a trajectory recovery model with kalman filter,which could predict the state of mis-object and associate it with detections in subsequent frames.The trajectory recovery model can reduce the fragments and improve the tracking accuracy.6.Experimental results show that the algorithm we proposed improved the tracking accuracy and the robustness to camera motion.Besides,the trajectory fragments also reduced.