Face Tracking under Varying Poses based on Particle Filter using a Combination of Adaptive Template and Background Detection

T.T. Son, H. Goto, and S. Mita (Japan)


Particle filter, face tracking, adaptive template matching, online learning, background detection.


This paper presents a method for solving face tracking problem under varying poses. The proposed method is based on Particle Filter (PF) using a combination of adaptive template and background detection. Difficul ties of face tracking problem are arbitrary movement and continuously-changing shape of facial objects. It is neces sary to select a suitable template online for a robust track ing. In our proposed method, the background detection uti lizes the online histogram at each position in each frame to classify pixels of moving object and background. Our method is independent of background change. It is usually slower than that of object movements. This background de tection helps to confine tracking regions in each frame for accuracy improvement. The adaptive template method uses an adequate template updated time by time for changes of facial pose. As a result, facial template can be adaptive to a rather large difference of facial pose. Our proposed method can run online with 5 frames/s and return a robust result of face positions in video sequences where facial ob ject changes from frontal to left, right, up, and down poses fast and continuously.

Important Links:

Go Back