Abstract:

To make full use of human face videos in temporal and spatial domains and obtain accurate face image sequence for face alignment, we propose a face detection framework in combination with a face tracking algorithm. We use the simple and fast frontal face detection algorithm to detect face images from videos, and then employ the detection results to initialize, check and adjust the face tracking results. However, the CAMShift tracking algorithm is vulnerable to the negative impact of skin color areas, which leads to more redundant information. To solve this problem, we propose an improved CAMShift-KLT algorithm, which utilizes interest points to trace the edge of the face image and acquires accurate face alignment images. Experimental results show that compared with the CAMShift algorithm, the proposed algorithm can obtain more accurate face regions, and meanwhile, it has smaller tracking migration distance, higher hit rate of tracking and tracking effectiveness. Besides, compared with the contrast algorithm, the improved  CAMShift-KLT algorithm can better track the face region.

Key words: face tracking, face detection, CAMShift, video, KLT