The project aims to build a multi-camera multi-target surveillance system to track people walking through the lobby.
First, we settle the environment setting, including camera cailibration , identifying and discretizing the ground plane in 3D coordinate.
After environment setting, we can project a cubic (used to approximate a person) located on the ground plane to image plane of each camera.
During tracking phase, we compute a probabilistic map representing whether a grid of ground is occupied by a person based on the image observations and build appearance models for each detected targets.
Then we can constructe the trajectories by connecting detected targets based on the probabilitic map and appearance models.
Figure 2. Some visualization results. For each image block, it contains two images captured at the same time from different perspectives.
Green boxes are the deteced targets and the corresponding numbers shown on the top of the boxes are the targets' ID.
F. Fleuret, J. Berclaz, R. Lengagne and P. Fua, Multi-Camera People Tracking with a Probabilistic Occupancy Map, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 30, Nr. 2, pp. 267 - 282, February 2008.