Two of the most pressing and related problems in computer vision research are motion analysis and the correspondence problem. To date, most of the studies deal with these two problems sequentially. The paradigm of conventional methods is a three-step model in which motion analysis (e.g., estimation of motion parameters and surface structure) rely on the result of the correspondence problem but not vice versa. One is therefore forced to resolve correspondence based on the properties of featured elements in successive image frames, but not on their geometric considerations which might allow a motion interpretation. If these two problems are treated on the same level simultaneously, then several advantages for understanding time-varying imagery could emerge. This research builds on previous work using this approach. The particular topics to be investigated include 1) correspondence and motion analysis in realistic models; 2) three or more view problems; 3) the use of line segments for motion estimation, and 4) motion analysis from perspective projection under restricted assumption.
