A region-based approach to face segmentation and tracking in video sequences

In this paper we propose an algorithm for face seg;mentation and tracking w hich is able to obtain the actual shape of a face present in a video sequen ce and track it in time. The face segmentation step relies on an initial pa rtition of the image, based on color homogeneity, which is fine enough to c ontain, among others, all the regions that form the face. Since a face cont ains a set of regions with chrominance homogeneity, the regions of the init ial partition are merged following this criterion. The sequence of mergings is represented as a binary tree. For each node in the tree a similarity me asure between the node and a face class is calculated. This measure is rela ted to the likelihood of an image of being a face. The node which maximizes this likelihood is selected, giving a first estimate of the face. The fina l face is obtained by a refinement step that merges the remaining face regi ons to the face estimate. In the tracking step, the face partition in the p revious image is motion compensated (projected) and the final face is obtai ned by a merging process.