Region-based segmentation and tracking of human faces

A new algorithm for face segmentation and tracking is presented. The face segmentation step relies on a fine segmentation of the image based on color homogeneity. In order to reduce the number of possible mergings, a merging order using chrominance homogeneity is applied. The sequence of mergings is represented as a binary tree. A set of regions (a node in the tree) is selected maximizing an estimation of the likelihood of being a face. The final face partition is obtained by a region merging process. The face tracking step uses motion information. The face partition in the previous partition is motion compensated (projected). The final face partition is obtained using the previous region merging process, driven by the projected face region.