Visual Tracking by Adaptive Kalman Filtering and Mean Shift

A method for object tracking combining the accuracy of mean shift with the robustness to occlusion of Kalman filtering is proposed. At first, an estimation of the object's position is obtained by the mean shift tracking algorithm and it is treated as the observation for a. Kalman filter. Moreover, we propose a dynamic scheme for the Kalman filter as the elements of its state matrix are updated on-line depending on a measure evaluating the quality of the observation. According to this measure, if the target is not occluded the observation contributes to the update equations of the Kalman filter state matrix. Otherwise, the observation is not taken into consideration. Experimental results show significant improvement with respect to the standard mean shift method both in terms of accuracy and execution time.