Motion vector search for 2.5D modeling of moving objects in a video scene

If we can estimate accurately the rotational and translational motion of a moving object in 2D images then we will be able to construct 3D model from 2D images, which is known as structure from motion (SFM). This paper modifies the MPEG like research technique in order to determine accurately the motion vectors associated with the moving objects in a video sequence. MPEG like search techniques applied to the RGB images does not necessarily mean geometrically similar points. The accuracy of the motion vectors is low in areas where the color appears uniform. We modified MPEG technique as follows. First, We employed exhaustive search to find where an object moved between two adjacent frames, the reference frame and the target frame using the minimum square difference (MSD) for the RGB images. Second for each frame we find the corresponding black and white (BW) image and we combine it with the gradient images. We apply the MSD to the combined images. Finally we combine the results of the MSD produced from the RGB images and the gradient and BW images. We found that the combined method achieved better motion prediction than either RGB alone or gradient alone