Object tracking integrating template matching and mean shift algorithm

Mean shift algorithm assumes the object motion is smooth with no abrupt changes, leading to failing to track the target when the target's speed is fast. Fast template-based tracking could tackle this problem but faces the difficulties like the loss of the optimal matching result and fixed-size template. We present a new object tracking approach integrating these two methods. Firstly, the template-based tracking method, which is speeded up by the coarse-to-fine strategy and dominant feature set, is employed to find roughly the candidates in the entire image. Then, a local mean-shift process is initialized in each candidate and these processes find the nearest local maximum in their respective neighbors. Among these local maxima, the position with the maximum one is regarded as the final estimate of object location. Finally, the target's current size is estimated and the template is updated accordingly. Experiments demonstrate the good performance of the proposed method.