Abstract

We propose a method of segregating moving objects which combines spatiotemporal filtering and feature matching. The matching scheme is preferred in many computer vision studies of shape from motion, since it produces accurate motion estimates and thus leads to sharp segregation. However, it has an inherent problem called false matching. To get around this problem, we combined the matching scheme with the spatiotemporal filtering scheme frequently used to model biological motion detectors. In our two-stage model, image sequences are first processed by spatiotemporal filters to obtain estimates of local motion. Several velocity candidates are detected in this filtering stage, and these are fed into the second matching stage to reduce the load of the matching process by limiting the range of correspondences. In the matching stage, local velocities are obtained for several kinds of feature. These features vote for velocity estimates, and the velocity with the highest score is assigned to each pixel as the best estimate. Moving objects are identified as sets of pixels with the same velocity. We used this two stage model in natural image sequences and obtained good results.

© 1989 Optical Society of America