A novel filter for block-based object motion estimation

Noises, in the form of false motion vectors, cannot be avoided while capturing block motion vectors using blockbased motion estimation techniques. Similar noises are further introduced when the technique of global motion compensation is applied to obtain “true” object motion from video sequences, where both the camera and object motions are present. We observe that the performance of the mean and the median filters in removing false motion vectors, for estimating “true” object motion, is not satisfactory, especially when the size of the object is significantly smaller than the scene. In this paper we introduce a novel filter, named as the Mean-AccumulatedThresholded (MAT) filter, in order to capture “true” object motion vectors from video sequences with or without the camera motion (zoom and/or pan). Experimental results on representative standard video sequences are included to establish the superiority of our filter compared with the traditional median and mean filters.