Abstract

We propose a new technique for tracking based on side statistics and maximum discriminant analysis. The object to be tracked is modeled by a set of sample points on the boundary together with image statistics inside the object. Tracking is conducted by maximizing the discrimination between the object and the background, based on an adapted Kullback-Leibler divergence without knowing the statistics of the background. Since no knowledge of the background is required, our technique is particularly useful in dynamic environments where the background can change substantially during the performance of a visual task, or when a system needs to be deployed in different environments. Because we use both the side statistics and the boundary information, our method is more robust than the traditional approaches that use either just boundaries or just regions. As will be shown experimentally, our technique can deal with complex environment, changing background, and partial occlusion, and it is real-time and accurate. We have used the technique to track a panel such as a piece of paper in an application system called VISUAL PANEL which serves as a wireless mobile input device to a computer.