Research On Robust Appearance And A Priori Constraint Based Object Tracking

As a hot topic in the computer vision field, the main purpose of visual object tracking is to determine the position of visual object in every video frame, which plays an important role in anglicizing and understanding the video content, such as automatic monitoring, human computer interaction, and video retrieval. However, in practice it is hard to track the visual object effectively when suffering various interferences, including the low resolution and noise of video capture device, shadow/illumination change in video scene, motion blur and occlusion of visual object and so on. Therefore, in this thesis we research on establishing a robust description model of visual object appearance, and combining a priori constraint in tracking process to improve the stability of tracking method in dealing with interferences. The detail contents of this thesis are as follows:(1) We introduce the background and significance of visual object tracking, and sort out the implementation process of it. Then we discuss the existing problems and corresponding causes in visual object tracking, and summarize the conditions a robust description model should have. Further more, we analyze the research status of visual object tracking, and also give a detailed description and evaluation of the specific implementation methods in each basic step of tracking process.(2) In order to track the object effectively, we research to express the object through regional nodes, which represent the different local regions of it, and then establish the structural relationship between regional nodes to complete the hierarchical structured description of visual object appearance. On that basis, we convert the tracking problem into estimating the motion states of regional nodes, and do the rough estimates through matching SIFT flow firstly, then by combining structured description, the roughly estimated results would be adjusted base on adaptive structure preserving constraint, which reduces the uncertainty caused by interferences and improves the precision of estimation to achieve robust object tracking finally.(3) For the purpose of maintaining the effectiveness of tracking method, and at the same time to simplify the visual object description and make the tracking operation easily, we research on using a simple chain-structure random field to express both the objects represented by different cues and the motion consistency priori constraints among them in a unified way. Then the tracking problem is converted to simply optimize the objective function of random filed, which not only considers each cue to have a high likelihood estimation of object position, but also requires the differences among estimated positions are small. Through this way, it can usually avoid the influence of abnormal situation and thus maintain the effectiveness of the tracking method.