Abstract
This thesis proposes an effective approach for kernel-based video object tracking. Video object tracking plays an important role in many applications of multimedia, such as video surveillance, object-based video database search, and automatic video manipulation. The main difficulty comes from the scaling change of a video object and occlusion. We formulate the tracking problem as a Bayesian learning framework and define the probabilistic models in terms of the distance between current frame and reference frame. We thus take the gradient method to obtain an optimized result with respect to a number of reference frames. Then we adopt the linear combination strategy and the modified decay rate to integrate the information from each reference frame and obtain the tracking result. In our experiments, we show the comparison of our method and the mean-shift algorithms. The tracking results of our method are more effective and reliable than the mean-shift algorithms. Even when the target object is severely occluded, our method tracks the reappeared object correctly.