Mean-shift Algorithms in Face-tracking

Choe Jeehyun

An approach to face-tracking in real-time video is proposed. Among many known methods, this experimental review specifically deals with the method using mean-shift algorithm. In this review, two experiments for the face tracking have been done including simple tracking and kernel-based tracking. First, the skin color modeling of the target area is done. Then the skin color filtering according to the model which has been calculated is processed through the whole image. Average location among the filtered pixels is calculated as a center of the tracking object. This is how the simple tracking is operated. The limitation of this algorithm is that when other objects with similar color to face exist in the background the tracker doesn't work well. Mean-shift algorithms supplement this problem. In mean-shift algorithms, the procedure only deals with the regions adjacent to the target spot on the prior frame and instead of dealing only with the average and the variance of the target color, a kernel-weighted histogram over the target area is used in a modeling procedure. While tracking a face, mean of the histogram is iteratively shifted and finally fixed to the target area. The experiment result shows that in the kernel-based tracking, more accurate estimation of the target area is performed than in the simple tracking. Even when the palm of the hand with the same skin color appears in the image, the tracker fixes to the face. The tracking is robust to the change of a background lights or colors compared with the simple tracking.