Tsinghua Face Detection and Tracking for CLEAR 2007 Evaluation

This paper presents the algorithm and evaluation results of a face detection and tracking system. A tree-structured multi-view face detector trained by Vector Boosting is used as the basic detection module. Once a new face are detected, a track is initialized and maintained by detection confidence and Lucas-Kanade features, which are fused by particle filter. Additionally, a post process is adopted to eliminate low confidence tracks and connect track fragments which are likely to belong to the same target. Evaluation results are given on video data of CLEAR 2007 test set.