Abstract

Motion estimation in image sequences is a fundamental issue in many applications as for instance in artificial vision and three-dimensional scene reconstruction. The basic problem of the recent spatio-temporal filtering techniques used in the perception of object motion from a time sequence of 2-D images is the computational burden. This is mainly due to the computation of the frequency responses of the images to the Gabor filters using the multi-resolution approach. In this paper we present a method for detecting and estimating object motion using spatio-temporal frequency information from image sequence. Our model of visual perception of object motion from a video stream is based on a contributory adaptation of spatiotemporal Gabor filters. This method uses a bank of Gabor filters which are frequency tuned and limited in spatial extent. Instead of testing the entire filter bank to determine the appropriate filter parameters, a genetic algorithm is used to derive the filter subset that provides the object texture information and optimize the search for its perceived motion. This model has been evaluated on artificial as well as natural image sequences. The results obtained show the feasibility of the approach which attains a reasonably fast output rate (several images/second) for a better resolution.

Key words: Dynamic vision, motion estimation, segmentation, genetic algorithms, Gabor filters.