The Research On The Robust Of Target Tracking In Complex Scenes

Target tracking is an important component of computer vision, given the target manually or automatically, tracking algorithms should meet the requirements of robust tracking in real-time. Real-time property asks for efficient searching algorithms. On the other hand, the tracking algorithm should be robust to the changes of motions, object gesture, and also the changes of the scenes. To enhance system robustness under the real-time requirement is not only one of the new and hot fields involved in the target tracking but also our goal in this thesis.Traditional tracking algorithms like Mean Shift based method could tracking object in real-time, but no consideration is given to occlusion, target location's accuracy reduces under occlusion. Corresponding solutions are proposed by many researchers, but these methods always need judge the beginning and ending of occlusion accurately. This thesis studies fragment tracking algorithm, which modeling target by multiple patches. By combining all patches' similarity at each candidate position to find the best state, this algorithm gets robustness target location under occlusion, overcomes the shortcoming of threshold difficult to choices in occlusion judgments. Then adopts integral histogram to reduce computational complexity of histogram, modifies matching technique to enhance the real-time performance. Based color contrast between object and surrounding region, choose patches better classified the object to improve the robustness under occlusion further.As traditional method used histogram, contour, template or other model relatively fixed, they can't adapt these video scene under scene dynamic change or object motion in large-scale. This thesis proposes a novel algorithm based on foreground feature points detection. Create background model from feature points, then classified foreground feature points by background model. By search matched points of target points in the global scope from foreground feature points, eliminates background feature points in the local scope, this method could reduce the impact of background feature points in the match process, and decrease interference by employing SIFT-like operator to descript feature points.We realize algorithm proposed in this thesis and other relative method programming by VC in PC platform to verify theoretical conclusion and the performance of algorithm. Experimental results indicate the proposed algorithm has good tracking performance under complex background like occlusion, scene dynamic change or object motion in large-scale. Excellent on keeping robustness under complex scene, this method has a low complexity, which can meet the requirement of real-time for higher layer vision analysis task. Therefore, the new algorithm is valuable in application.