Maximum Likelihood Estimation of the Template of a Rigid Moving Object

Motion segmentation methods often fail to detect the motions of low textured regions. We develop an algorithm for segmentation of low textured moving objects. While usually current motion segmentation methods use only two or three consecutive images our method refines the shape of the moving object by processing successively the new frames as they become available. We formulate the segmentation as a parameter estimation problem. The images in the sequence are modeled taking into account the rigidity of the moving object and the occlusion of the background by the moving object. The segmentation algorithm is derived as a computationally simple approximation to the Maximum Likelihood estimate of the parameters involved in the image sequence model: the motions, the template of the moving object, its intensity levels, and the intensity levels of the background pixels. We describe experiments that demonstrate the good performance of our algorithm.