Research On Face Detection And Tracking Algorithm Based On Video

Recently, human face analysis has been a hot topic in computational vision and patternrecognition field as the developing of information technology. It has essential research valuein identification, video surveillance, intelligent human-machine interface system and so on.The research of human face analysis involves many subjects, such as detecting, tracking,identification, expression analysis, animation, modeling, and etc…A stable and fair detectionand tracking algorithm is the basis and guarantee for other human face analysis technologies.Although video-based face detection and tracking has been widely researched and manyeffective algorithms have been proposed, there are still a lot of difficulties in developing arapid accurate and stable face detection and tracking system when there is so muchinterference such as illumination variation, similar background, attitude change, occlusion, fastmotion and so on.This paper study on video-based face detection and face tracking, and these two parts caninfluence each other. In the aspect of face detection, considering the influence of complexfactors in video, the face detection algorithm is divided into prediction and recognition twosteps. First, Reference White algorithm is used to compensate illumination and clear the colordeviation in the images. Second, YCb’Cr’ color space is chosen for skin color modeling as itsillumination robustness and well performance in skin color clustering, so that skin color regionis segmented as predicted area. Then AdaBoost algorithm is used to recognize faces based onthe fetched skin color area. The joining of multi-view face training samples to the originalones makes the trained classifier has a better recognition rate of multi-view faces. This methodimproves detection rate and has a better detecting speed, and satisfies the real-timerequirement of following tracking.In the aspect of face tracking, traditional Camshift algorithm has disadvantages likesensitivity of illumination and alike background, and can not recover target after losing. Forresisting the interference of skin color alike area, the proposed algorithm considers thebackground information into Camshift framework to assign weight of different parts of facedynamically. In the tracking process, face is divided to keep geometric structure at some level,and the location of face is more accurate when occlusion happens. Besides, based on traditional SURF feature point tracking algorithm, random ferns is added to filtrate candidateregion and shortens the time of SURF feature point matching. Finally, accelerated SURF iscombined with improved Camshift and the problem of Camshift losing target is overcome.Theexperimental results shows that it will improve the performance of tracking and stability of thesystem in real-time demand.