The movies on this page are sample output from our system. All the movies were created in real time, with images taken at 15 fps. We tested our program on three cases. One was a single object tracking test. It was performed without any noise or obstacles. This also provides a good demonstration of how the disparity map is used in calculating the snakes. The snake tracked the object perfectly even when the shape of the object was altered in a variety of ways.

The Chair

The second test was performed on a still object and a moving object. We tried a variety of still objects, for example the chair in the image. When a person passed by behind the chair, the person was partially occluded and we could restore his boundary correctly. However when the person passed in front of the chair, the person completely occluded the chair and sometimes the boundary of the chair could not be restored. This also sometimes led to losing the person's boundary. Our experiments with other objects led to similar results. In approximately 95% of cases of partial occlusion, the boundaries were perfectly restored. Our system recovered from complete occlusion much more rarely, about 30% of the time.

Real World

The third type of test performed was between moving and interacting objects. In this case, we limited the testing to partially occluded subjects. The boundaries remained stable when the subjects moved horizontally and diagonally. However, when people were moving vertically (towards the camera), and their legs were occluded, the boundary often rose over the legs. This is reasonable, due to the visual growth of the subject as he approaches the camera. This caused our algorithm to guess the occluded points inaccurately. As a result, the legs of the person were engulfed by the occluding person or object.

Overall, though our system may not be ready for the real world, we appear to have had substantial results in tracking people and objects in real-time through the given environment. The results were excellent for horizontal and diagonal motion. We feel that further work on refining the algorithm would yield the necessary steps to make it a fully stable tracking system.