Head segmentation and tracking play an important role in computer vision and digital media. Due to the complexity in real scenes, face region and background undergo high diversity of visual appearances. Thus, it has been an important task in computer vision to design robust, stable and adaptive head segmentation and tracking algorithms. From the aspect of improving the generalization ability and discrimination ability, this thesis focuses on the relevant problems in head segmentation and tracking. The main contributions of the thesis are listed as following: 1、We propose a discriminative color prior model, and apply it in feature matching based 3D head tracking. Based on the property that the face region usually owns different colors with background, the model classifies feature points into two kinds and identifies the outlier points in background. Rejecting the outliers, the wrong correspondences in feature point matching will be reduced, hence the iteration times and computation time in pose estimation will be decreased. Experimental results show that it owns sufficient discriminative ability and it could improve the head tracking algorithm to be more accurate and robust. 2、We propose a skin detection algorithm based on linear regression tree. Its central idea is to utilize the tree classifier to partition the color space into several sub-regions and to define a decision function between skin and non-skin in each sub-region, which is suitable for the multi-mode property of skin colors. The tree is fast to train, fast to test and it does not call for a large training set. Experiment results on public skin database shows that it owns both good generalization ability and discriminative ability, while the tests in real videos further imply its efficiency. 3、An automatic head segmentation method is proposed based on semi-supervised classification. It is modeled in the semi-supervised classification framework with local spline regression based graph regularization. With the automatic supply of supervised information, it works without human interactions. To enable the correctness of supervised information, an initial head region is estimated and utilized to constraint them. Experimental results show that it could obtain comparative results with interactive methods and it is robust to head pose variations, scale changes and different lighting conditions. 4、We propose an integrated head segmentation and tracking method. It combines head segmentation and ...
修改评论