The use of model-guided grouping in model-based motion tracking

A general motion tracking system consists of three major parts: segmentation, correspondence and viewpoint verification. This thesis presents improved methods for solving the correspondence problem. Most previous approaches to solving the correspondence problem have used lower-level primitives such as points and lines as their matching tokens. It has been found that if the matching tokens, such as points and lines, are less distinctive, greater ambiguity in matching inevitably arises. The objective of this thesis is to develop a robust solution to the correspondence problem even when the frame-to-frame motion is relatively fast. A new approach called Model-Guided Grouping, which is used to derive intermediate-level structures as our matching tokens, has been introduced. The term Model-Guided comes from the fact that the groupings are guided and derived locally, with the contemporary use of model structures, around the predicted model during the object tracking. We choose junctions and parallel pairs as our matching tokens, thus the information coded in these structures is relatively invariant in consecutive frames. The matching strategy is coarse-to-fine, and partial matching will also be allowed when occlusions are present. The method for evaluation of probability of accidental match based on junction groupings will be discussed. Comparisons are made between the current system and Lowe's previous system. The results show that with a slight increase in computational cost, the use of the higher-level groupings as matching tokens leads to a much more robust model-based motion tracking system. Keywords: motion tracking, perceptual grouping, model-guided grouping, correspondence, model-based vision, segmentation, token matching, viewpoint verification.Science, Faculty ofComputer Science, Department ofGraduat