Mean shift blob tracking with kernel histogram filtering and hypothesis testing

We propose a new adaptive model update mechanism for the real-time mean shift blob tracking. Since the Kalman filter has been used mainly for smoothing the object trajectory in the tracking system, it is novel for us to use adaptive Kalman filters for filtering object kernel histogram so as to obtain the optimal estimate of object model. The acceptance of the object estimate for the next frame tracking is determined by a robust criterion, i.e. the result of hypothesis testing with the samples from the filtering residuals. Therefore, the tracker can not only update object model in time but also handle severe occlusion and dramatic appearance changes to avoid over model update. We have applied the proposed method to track real object under the changes of scale and appearance with encouraging results.