Research Of Robust Object Tracking Based On Spatio-Temporal Model

With the rapid development of computer science and artificial intelligence, object tracking has gradually become one of the most important research areas in the field of computer vision. Object tracking is usually regarded as a process that the tracking method can robustly locate the target in the subsequent video in real-time after the target has been labeled manually. A robust tracking method can locate the target accurately even the target suffers from motion blur, complex background, illumination changes and scale changes. Therefore the study of strong robust and real-time tracking method is always a hot research topic, which is also the purpose of this thesis.This thesis mainly studies the tracking method based on the spatial-temporal model via the target and its around information. By using spatial-temporal model to construct the search area and refining the updating strategy of the spatial-temporal model, this thesis has proposed two new tracking methods. The main studying contents and results are summarized as follows:1. In the process of tracking, the information around the target is also very important, which can be used to locate the tracked target. This thesis analyzes the relationship between the target and its surrounding information which is used to build spatial model. The relationship between the adjacent frames is used to build spatial-temporal model. The parameters of shape and learning rate are discussed thoroughly which help us to gain insight into the robustness of spatial-temporal model. This part of work laid the theoretical foundations for the subsequent chapters.2. The traditional compression method extracts samples around the previous target region within a fixed search radius. The method is obviously not reasonable. When experiencing sudden acceleration motion, the tracker is easy to lose the target. Once the target is lost, it is hard for the tracker to find the target again. To solve this problem, we propose to use the spatial context between the target and its surroundings. Moreover, the spatial-temporal model is also constructed by the temporal relationships between successive frames. The spatial-temporal information can be used to calculate the confident map of the probability for the target location. The region with high confidence suggests the high possibility that target exists. Thus the samples can be extracted in the high confidence area. Then, the optimal target location can be estimated with a naive Bayes classifier using sparse coding features. Experiments show that using the spatial-temporal information to build the scope of the candidate region can better constraint the range of the target, which can make the method more accurate and robust.3. When the scale of the target is obviously changed in the process of tracking, the spatial area cannot be effectively updated. Thus the spatial-temporal model cannot well adapt to the scale change and the tracker will lost the target. According to this problem, we propose to construct the history target template library using cluster theory. We choose the object sample which can optimally represent the object state to form the template library. The multi-scale rectangular samples are extracted according to the center of the target estimated by the spatialtemporal model. The similarity between templates and samples can be analyzed by the gradient histogram feature. We choose the optimal scale and the corresponding sample to update the spatial-temporal model. Thus the target with scale change can be successfully tracked. Experiments show that our tracking method can efficiently adjust to the scale change. The spatial-temporal model can be updated in the condition of scale change, partial occlusion, illumination change, motion blur and some other challenges.4. In order to verify the effectiveness of the algorithm and the feasibility of the theory, experiments upon a large number of video images are performed. The results show that the proposed methods in this thesis can not only effectively adjust to the interferences of scale change, illumination change, complex background, similar object, etc., but also track the target in real-time.