Estimating 3D camera motion without correspondences using a search for the best structure

A novel approach is presented for recovering the motion parameters of a camera from two frames. The proposed method does not require establishing point correspondences between the images, as does most current techniques. Our approach is also more straightforward than the very few non-correspondence motion estimation algorithms. It is based on the estimation of structure for each given set of motion parameters. This resulting structure is then evaluated in an optimization process using saliency metrics, until the best structure and motion parameters are obtained. In this work we have devised and tested two different structure metrics: the first based on scatter and the second using tensor voting. Experimental results show that this method is effective and can be used in video-based scene modeling systems.