Multiresolution estimation of 3-D motion parameters

We propose a multiresolution solution to the problem of motion estimation where objects in the world may move relative to each other and the camera. Our basic contribution is twofold. Firstly, we show that contrary to current assumptions, the error in estimate of time-derivative is not zero-mean IID Gaussian. We show that it is image dependent and use wavelets to estimate its mean and covariance. Experimental results show a significant improvement in motion estimates using our technique. Secondly, we formulate a multiresolution solution to the (as yet unsolved) object motion problem. We assume an a-priori probability density of motion and our method works best when motion is small, as is the case in a sequence of television images. Our technique is distinct from wavelet based pyramidal searches