Multi-histogram Based Face Tracking Using Particle Filter Embedded in Incremental Linear Discriminant Model

This paper presents a multi-feature integrated algorithm which incorporates the particle filter and LDA model for face tracking. To solve this drift problem, the discriminant model for each feature is built up by considering face and background information to separate the face from background clutter. Similar to Adaboost method, we determine the reliability of each feature by the proposed measuring method, which is used for successive calculation of observation probability in a particle filter. Moreover, to address the face appearance variations and background changes, the Linear Discriminant Analysis (LDA) model for each feature is updated by face data which is selected according to the co-training concept. Experimental results showed the proposed multi-feature integrated algorithm is able to handle face appearance variations including out-of-plane rotation, partial occlusions, varying illuminations, scale or viewpoint changes, and mess background.