Learning distribution metric for object contour tracking

A new approach to tracking using active contour model is presented. Suppose that the class of objects to be tracked is characterized by a probability distribution, we tackle the active contour tracking problem by learning a suitable distance measure between distributions. A cross bin criterion for comparing distributions in quadratic form is adopted in this paper for active contour tracking, in which the measure matrix is learned and updated on-the-fly based on convex optimization. We model the image energy by the distance between the foreground distribution and the model one, divided by the distance between the background distribution and the model one. The experimental results have demonstrated the effectiveness and robustness of our method.