A GM-based multi-layer method for object tracking in video sequences

This paper introduces a novel method for object tracking. In the classification perspective, object tracking seeks the separation of two classes: foreground, which is the object, and background. Our method divides each frame into small blocks of uniform size, and a block is considered to be an object block, if it contains any part of the object, or background block otherwise. Generalization machine (GM), a newly developed learning tool, is employed to accomplish the classification task. After learning through a manually labeled training frame, GM yields a decision function which is used to classify blocks in each subsequent frame. The identified object blocks in the same frame builds up the whole object of interest. To deal with large and complex objects, a multilayer approach is proposed to avoid the complexity of nonlinear mapping, and to improve performance. Experimental results verify that the new method is fast and robust.