engleski

Establishing Motion Correspondence for Moving Non-rigid Objects

Here we present some well known motion correspondence algorithms applied to tracking of multiple similar moving non-rigid objects. Establishing motion correspondence is treated as a global optimization problem. Typically, a cost function that includes speed, direction or positional terms is optimized over all possible mappings or by a greedy algorithm either over a pair of subsequent images o r in an extended sequence. In our experiments tracked objects are not distinguished by some strong feature, they are able to move fast and in an unpredictable manner and they touch and even occlude each other frequently. In the presence of contact, when feature points do not represent object centers of gravity, correspondence algorithms degrade in performance. By a heuristic extrapolation of object positions in the case of the contact, we were able to substantially improve motion correspondence without sacrificing ability of our system to analyze all relevant data from a lengthy image sequence in real-time on a commodity computer.

computer vision; motion correspondence; multiple object tracking

nije evidentirano