| Tracking multiple people in single-camera video requires solving the occlusion problem, which occurs when the object being tracked is partially or fully invisible for some frames. We approach this as a maximum a posteriori probability (MAP) estimation problem, and show that appropriate choice of color and spatial pdfs makes it possible to establish correspondences between objects over time, even under occlusion. We have shown that a uniformly contaminated normal-kernel distribution is appropriate for modeling the spatial cue in occlusion scenarios. We demonstrate results on several different types of example sequences, containing 100% occlusion, object handover from one person to another and velocity reversal during occlusion.; If additional cameras are added to a tracking system, the need to establish correspondence between views of the same person seen in several cameras arises. We call this the consistent-labeling problem, and formulate it using two correspondence layers, one at the single-camera level, and the other at the multiple-camera level. Our framework, based on the extraction of Field-of-View lines, automatically discovers the spatial relationships between cameras, and is simpler than competing approaches. We present two schemes for automatic initialization, depending upon the state of the environment. The homography between all cameras is recovered efficiently through either of these methods. Such a system is useful in many applications; in particular, for reorganization of video streams from camera-centric to object-centric, for generating global environment maps and for occlusion resolution.; Finally, we generalize the tracking problem to the video segmentation problem, which we view as tracking of all objects in an image. We show that these two problems are very closely related to each other. We present a framework for using multiple cues in MAP, and advocate the use of Logarithmic Opinion Pooling (LogOP). We demonstrate results on complex video sequences consisting of several hundred frames. Our segmentation results are very accurate, and combine the strengths of motion segmentation and color segmentation together in one framework. Resulting segmentation can be used for video interpretation and MPEG4-type compression. (Abstract shortened by UMI.)... |
