Camshift tracker based on multiple color distribution models

Traditional Camshift tracker based on a single color histogram model is not robust to appearance changes of the target caused by changing viewpoint. To tackle the problem, a possible way is to use a model with more powerful representation ability. In this paper, we model the target with multiple color distributions according to prior knowledge of the target and then design a cost function. Through minimizing the cost function, the optimal model is selected in real time from the convex combination of model sets for tracking in the next frame. In addition, when researching Camshift tracker in detail, we find the relationship between the average intensity of probability image and the color distribution histogram of image pitches, which helps to illuminate the mechanism of model selecting process. Experimental results conducted on head sequences demonstrate our tracker can deal with dramatic appearance changes of target in an elegant manner with low computational cost when compared with Camshift tracker with a single fixed model or single adaptive model.