Research On Key Technologies Of Multiple Object Tracking Based On Data Association

With the rapid growth of intelligent video analysis,multiple object tracking(MOT)plays an increasingly important role in traffic monitoring,traffic statistics and other computer vision.MOT is aimed at locating objects with significant characteristics in the video sequence and assigning each object a unique identity to predict their trajectories.However,duo to the complexity of the tracking background,the deformation of objects,the similar appearance between objects and occlusions,there are great difficulties in MOT.In recent years,due to the significant improvement of the object detection algorithm,a tracking algorithm based on data association has been developed.The tracking algorithm and the detection algorithm are combined to link the detection results to the tracking process and form a trajectory.Although this approach largely solves the problem of deformation and partial occlusion in the tracking process,it still can not deal with objects with similar appearance or severe long-term occlusion.In this thesis,after studied the data association algorithm,feature expression of objects and occlusion processing algorithm in-depth,a new framework of MOT system is proposed to improve the tracking accuracy and reduce the computational complexity.The main research contents include:1.Choose the Deformable Part-based Model after a brief overview of the multiple object detection algorithm,which can provide more accurate input for the tracking system.2.Summarize and compare several common data association algorithms,in which the network flow model has better performance and is more suitable for the complex scene,which is the focus of this research.3.Study the minimum cost network flow algorithm.It is found that the complexity of the dynamic programming algorithm is lower,so the data association can be realized in less time.4.Study the multiple object feature expression method,and propose the method based on the combination of global feature and local feature,which effectively solves the problem of identity switch and false alarm in the tracking.5.Do the research on occlusion problems,the Extended Kalman Filter-Explicit Occlusion Model(EKF-EOM)is proposed to deal with the trajectory interruption and easily lost objects.The innovation of this paper is as follows:(1)in the network flow model,the combination method of global feature and local feature is proposed to express the characteristics of objects,which is easy to distinguish similar appearance between objects.(2)an improved occlusion model is proposed,which aims at estimating the information of objects that are lost due to occlusion by establishing an explicit hypothesis based on Extended Kalman Filter.