Abstract
In this paper, we consider the problem of estimating the instant motion parameters for any image among the image sequences and its application to motion-based segmentation problem. The parameters we estimate are the location of edges and their corresponding velocities and orientations. One of the most difficult problems in motion estimation and motion-based segmentation is to tackle the ”chicken-and-egg problem” considered earlier in [1], wherein one needs to proceed back and forth between motion-based estimation and motion-based segmentation. An important contribution of this paper is the development of an algorithm which can estimate motion parameters along moving discontinuities, edges and boundaries in a given sequence of images, using velocity tuned filters. Such an algorithm provides motion parameter estimates and edge locations, both simultaneously without any need for iteration. Also the proposed algorithm can be generally applied to many types of motion, not only restricted to translational motion. It does not need point correspondence or any stochastic models. We use velocity tuned 2D＋T filters to perform motion estimate. After we simultaneously obtain the edge locations and motion parameters, we can further apply them to motion-based segmentation using clustering method. A ridge-skeleton on the reconstructed edges in each cluster provides the final motion-based segments.The algorithm has been simulated and tested on real image sequences acquired by a moving camera and the result has been found to be satisfactory.
