视频中多线索的人脸特征检测与跟踪

Multi-Clue Based Facial Feature Detection and Tracking in Video

针对目前的人脸特征检测与跟踪算法存在的对环境适应能力差、缺乏自我检错能力的缺点 ,该文提出了一种多线索综合的新方法 .多线索中包括基于深度信息的人脸区域粗分割、基于多关联模板匹配的人脸检测、利用多尺度Sobel卷积的特征提取、基于“特征眼”的人眼验证以及基于多视图的校验方法 .多种线索互相补充、自我检错和纠错 ,对背景、光照及姿态变化具有较强的适应能力 .实验表明该方法是有效的、鲁棒的 .

Multi-clues include rough face region segmentation based on disparity or color information, face detection based on multiple related templates matching, feature detection based on multi-scale Sobel convolution, eye feature verification based on eigen-eyes, and facial feature verification with both geometry and rigid plane motion constrains in multiple views. First, binocular stereo video input is used for robustly extracting head region from complex back-ground through disparity clustering. Then, the multiple related template matching method is applied to find the accurate face region from this rough segmentation. Facial organ candidates are extracted from the detected face region at a specific scale space called organ scale for Sobel filter. Eye pair is chosen from candidates by eigen-eyes method. Finally, nose and mouth corners are detected according to projections. The algorithm can automatically switch between facial feature detection and tracking based on embedded verification procedure. In this method multiple clues are joined together to supplement each other, which makes automatically error-checking and even error-correcting possible, which greatly improves the algorithm's adaptability to lighting and face pose changes under complex background. Experiment results over 189 video sequences demonstrate its effectiveness and robustness.