Moving Object Tracking Modeling Based On Manifold Learning

As one of the most active research in the field of computer vision, image processing,pattern recognition, moving object tracking technology has been playing an extremely important role in security surveillance, military affairs, human-computer interaction and many other areas. Although researchers have proposed many effective object tracking methods, the tracking performance is not stable and robust because these methods track the object in image space directly and thus the dimension is very high. Therefore, in this thesis, the manifold learning method is introduced into object detection and tracking in order to obtain stable tracking performance.The main contributions of this thesis are as follows:1. A multi-feature fusion method is proposed for target modeling. Five uniform modes related to the edge and the corner of the texture models of Locally Binary Pattern(LBP)are extracted first and then fused with the color feature for modeling the target, which can increase the observed information of the target and improve the robustness of the observation model.2. The head pose estimation model is built based on manifold learning. Through comparing different manifold methods for head poses estimation, we conclude that the Locality Preserving Projections(LPP) algorithm based on color-texture feature is the optimal.It can describe the movement of the head more accurately.3. Moving object tracking based on manifold learning. A framework of moving object tracking based on manifold learning is proposed. Color- texture similarity comparison are used to get the better candidate area.Then the LPP algorithm was applied to tracking. It can track on the low space directly after the dimension reduction. Experimental results show that the proposedalgorithm is feasible and has high accuracy and robustness during object tracking.