Tracking of people in crowded scenes is challenging because pecause people occlude each other when they walk around. The latest revision of our person tracker uses adaptive appearance models that explicitly account for the probability that a person may be partially occluded. All potentially occluding targets are tracked jointly, and the most likely visibility order is estimated (so we know the probability that person A is occluding person B). Target size adaptation is performed using calibration information about the camera, and the reported target positions are in real-world coordinates.

This movie shows real-time tracking of people under occlusion. Note that people are succesfully tracked even when they walk in front of each other, or when they walk very close to each other and then retreat. (6.5MB)

This movie shows real-time servoing using a camera mounted on a Pan-Tilt head. The camera is able to successfully track a fast moving object of interest. (4.8MB)

This movie show real-time tracking of a person in two cameras. The top figure is a top-view showing the estimated world location of the person (marked by X). (2.7MB)