|
|
 |
|
|
| PeopleVision | |||
|
|||
|
|
2D TrackingThe 2D tracking relies on appearance models which are image templates. New appearance models are created when an object enters a scene. In every new frame, each of the existing tracks is used to try to explain the foreground pixels. The fitting mechanism used is correlation, implemented as minimization of sum of absolute pixel differences over a predefined search area. During occlusions, foreground pixels may be overlapped by several appearance models. Color similarity is used to determine which model lies in front and infer a relative depth ordering for the tracks. Once this relative depth ordering is established, the tracks are correlated in order of depth. The correlation process is gated by the explanation map which holds at each pixel the identities of the tracks explaining the pixels. Thus foreground pixels that have already been explained by a track do not participate in the correlation process with more distant models. The explanation map is now used to update the appearance models of each of the existing tracks. Regions of foreground pixels that are not explained by existing tracks are candidates for new tracks. A detailed discussion of the 2D tracking algorithm can be found in [1, 2]. The 2D tracker is capable of tracking multiple objects moving within the field of view of the camera, while maintaining an accurate model of the shape and color of the object.PETS refers to data from the IEEE Performance Evaluation of Tracking and Surveillance workshops. All demo videos are in MPEG1 format. 1. PETS 2001 surveillance type tracking (video13.5MB) 2. PETS 2002 tracking people infront of a store (video
8.8MB) [1]
Appearance Models for Occlusion Handling [2]
Tracking with Probabilistic Appearance Models Other Research Areas:
|
|||||||||||||||
| About IBM | Privacy | Legal | Contact |
Other
projects in
