Real-time person tracking with an active camera should be fast and robust because the current camera control affects the tracking results in the successive fames. There are various algorithms for target tracking such as motion-based, shape-based, and color-based methods. A motion-based method is based on target motion(s) different from the background motion. But for the cases of severe pan-tilt rotation and zoom control, it is difficult to distinguish the target motion from the background motion. In addition, since it can only detect moving parts of the target, its scale information is not always accurate. A shape-based method can provide the reliable information of the scale of target, by detecting target boundaries. But its result can be easily affected from clutters in the background. A color-based method, especially the one using a color histogram, provides a quite robust tracking result for a non-rigid moving object. However, if the target histogram varies and a similar color exists near the target, it selects a position quite far from the target center, resulting in a wrong scale.
In this thesis, we propose a robust and fast head tracking method, which is composed of a color-based position search and a shape-based refinement. The color-based position search selects a position near the center position of the target head using the model color histogram and the temporal color consistency with the previous result. Meanwhile, the shape-based refinement step offers the correct information regarding scale and position.
A real-time tracking is implemented on the basis of the proposed method by using a PC of 2.4 GHz CPU, CCD camera, and pan-tilt unit. The experimental results show that the system can provide very robust tracking results with the frame rate of 40fps.