Tracking the Human Body Using Multiple Predictors

Abstract

The objective of this work is to track the human body from a video sequence, assuming that the motion direction is parallel to the image plane. Tracking the human body is a difficult task because the human body may have unpredictable movements and it is difficult to accurately detect anatomic points in images without using artificial marks. Furthermore, self occlusions often prevent the observation of some body segments. This paper describes a tracking algorithm, which avoids the use of artificial marks. The proposed system is able to learn from previous experience, and therefore its performance improves during the tracking operation. The ability of the tracking system to gradually adapt to a particular type of human motion is obtained by using on-line learning methods based on multi-predictors. These predictors are updated in a supervised way using information provided by a human operator. Typically, the human operator corrects the model estimates several times during the first few seconds, but the corrections rate decreases as time goes by. Experimental results are presented in the paper to illustrate the performance of the proposed tracking system.