Object Tracking is an important computer vision problem and has been investigated during the past decades. It has wide application in human-computer interaction, intelligent surveillance, medical image analysis, robot navigation and video sequence analysis etc. In this dissertation, we focus on kernel tracking and silhouette tracking. Two new specific tracking algorithms are proposed for different applications. Our work can be concluded as follows: (1) For the general object tracking, a new hybrid kernel tracking algorithm is proposed, which combines discriminative and descriptive methods of object representation. In this method, object tracking is treated as classification problem of image pathes, multiple scale patches are used to model both object and background, the discriminative 2 class SVM and descripitive 1 class SVM corporate together to reduce the affect of “model shift” on tracking. The extensive experiment results show our algorithm’s superior to the state-of-the art algorithms. (2) For articulated pose estimation, an interactive tracking method based on image manifold and motion constrains is proposed. Compared with other generative methods, our algorithm take advantage of the combination of multiple motion constrains which greately reduce the computation complexity. The following experiment shows that our tracker can successfully estimate the pose state of human body.
修改评论