Point matching in a time sequence of stereo image pairs and its parallel implementation on a multiprocessor

An algorithm for finding point correspondences among stereo image pairs at two consecutive time instants (t/sub i-1/,t/sub i/) and its parallel implementation on an Intel ipsc/2 hypercube multiprocessor system are presented. There are 137 unambiguous matched-point pairs among the images used. This number of points may be not enough for generating the structure of the object in the images; however, for motion estimation using stereo imagery, only three pairs of points (for the ideal case) are required, which implies that the results are more than enough for the estimation of motion parameters. From the parallel implementation of the feature extraction and stereo match algorithms on a hypercube multiprocessor system, it is observed that if the computation is uniformly distributed across the image (feature extraction) then almost linear speed-ups can be obtained by partitioning the data equally among the processor of the hypercube.<<ETX>>