Mean Shift Tracking Algorithm with Scale Adaptation

This paper presents an adaptive window object tracking method for Mean Shift based on graph cuts theory. It copes with the size-changing object during visual tracking while the traditional Mean Shift can't change the scale of tracking window in real time. According to the Mean Shift iteration result of every frame, graph is created by using skin color Gaussian mixture model in a small area around it. Graph cut is implemented by calculating the minimum energy function based on max flow/min cut principle. And then the largest skin lump is found, which is accepted as tracking object in the result of graph cuts. As a result, tracking window size can be updated by the largest skin lump. Experimental results clearly demonstrate that the method can avoid the problem of nonstop shrinking bandwidth effectively which is brought by expanding and shrinking 10% of kernel function bandwidth. It reflects the real scale change of tracking target in real time, avoids the interference of other targets in the background, and has good usability and robustness. Besides it can be applied to controlling entertainment games that enriches operation mode of human computer interaction.