Robust multiple face tracking via mixture model  

作　　者：郭超 

机构地区：[1]Institution of Information Science and Communication Technology,Zhejiang University

出　　处：《Journal of Harbin Institute of Technology(New Series)》2010年第6期830-836,共7页哈尔滨工业大学学报（英文版）

摘　　要：Based on particle filter framework,a robust tracker is proposed for tracking multiple faces of people moving in a scene.Although most existing algorithms are able to track human face well in controlled environments,they usually fail when human face appearance changes significantly or it is sheltered.To solve this problem,we propose a method using color,contour and texture information of human face together for tracking.Firstly,we use the color and contour model to track human faces in initial images and extract pixels belonging to human face color.Then these pixels are used to form a training set for setting up texture model on eigenspace representations.The two models then work together in following tracking.To reflect changes in human face appearance,update methods are also proposed for the two models including an adaptive-factor by which the texture model can be updated much more effectively when the human face's texture or rotation changes dramatically.Experiment results show that the proposed method is able to track human faces well under large appearance rotation changes,as well as in case of total occlusion by similar color objects.Based on particle filter framework, a robust tracker is proposed for tracking multiple faces of people moving in a scene. Although most existing algorithms are able to track human face well in controlled environments, they usually fail when human face appearance changes significantly or it is sheltered. To solve this problem, we propose a method using color, contour and texture information of human face together for tracking. Firstly, we use the color and contour model to track human faces in initial images and extract pixels belonging to human face color. Then these pixels are used to form a training set for setting up texture model on eigenspace representations. The two models then work together in following tracking. To reflect changes in human face ap- pearance, update methods are also proposed for the two models including an adaptive-factor by which the texture model can be updated much more effectively when the human face＇ s texture or rotation changes dramatically. Experiment results show that the proposed method is able to track human faces well under large appearance rotation changes, as well as in case of total occlusion by similar color objects.

关 键 词：face tracking OCCLUSION EIGENSPACE eigenbasis particle filter 

分 类 号：TN91[电子电信—通信与信息系统]