Parallel distributed detection of feature trajectories in multiple discontinuous motion image sequences

This paper is concerned with three-dimensional interpretation of image sequences showing multiple objects in motion. Each object is assumed to have point features which are detected automatically and exhibits smooth motion except at times when a motion discontinuity may occur. The problem of estimating initial feature trajectories is that of matching between the first two frames. As more images are acquired, existing trajectories are extended. Both initial detection and extension are formulated as energy minimization problems that are solved using a Hopfield network. The energy function incorporates pertinent 2-D and 3-D geometric constraints. A second, 1-D Hopfield-like network is used to eliminate wrong matches among those found by the first network.