Multi-Object Tracking Based On Appearance Model And Interaction Information

With the explosion of video data in modern society,the manual way of data pro-cessing can no longer be a practical solution.It is an sianificant research direction in the field of computer vision that how to get the needed information efficiently and accu-rately from the video data.Visual tracking,especially multi-object tracking has become an important and basic research direction among these tasks.In the case of public secu-rity,intelligent transportation and other video data processing situations,we need to get the trajectory of each target,and then determine whether there is an anomaly happening or a traffic violation.In addition,multi-object tracking has been extensively used in the field of human-computer interaction,automatic driving and so on.In the tracking of multi-objects,there is a variety of complex situations such as oc-clusion,illumination changes.It is hard to distinguish different targets and track them.The most commonly used method of multi-target tracking is to detect firstly and then track them.After getting the detection responses,a tracker is used to associate these de-tection responses to trajectories of different targets.While tracking,visual information that is to say features is necessary.The requirement of these features is that they can keep the appearance consistency within track and the discrimination between tracks.It is hard for common features to satisfy all these 2 requirements.While tracking a tar-get,occlusion often occurs,so that the visual features of targets can not be obtained or the appearance of the target changes,so tracking rely solely on visual features fails.While tracking pedestrians,the motion model of multiple targets is influenced by each other,the motion models of the targets around contribute to the tracking of the cur-rent target.The utilization of interaction information of target can improve the tracking performance.The research contents and innovations of this dissertation are mainly the following two points:1.We proposed a real time multi-target tracking algorithm based on global and lo-cal features.In which two steps are presented which is based on global feature and local feature.The commonly used color histogram as a global feature is utilized among the global feature tracking,satisfying the requirement of consis-tency.Among local feature tracking,a new local feature based on Maximally Stable Extremal Regions is utilized to satisfy the requirement of discimination.Thus the requirement of consistency and discrimination can both be satisfied,and the computational complexity is low.So can be applied to real-time tracking.2.We proposed a multi-target tracking algorithm based on formation stability.On the basis of F-Formation,the concept of formation stability is put forward,and utilized to improve the performance of multi-target tracking in semi-crowded en-vironment.While tracking,in addition to the use of traditional visual information,the stability of the whole formation need to be maintained after association of two different tracklets.so that formation formed by pedestrians do not change greatly and the performance of tracking is improved.