Multiple video cameras are cheaply installed overlooking an area of interest. While computerized single-camera
tracking is well-developed, multiple-camera tracking is a relatively new problem. The main multi-camera
problem is to give the same tracking label to all projections of a real-world target. This is called the
consistent labelling problem.
Khan and Shah (2003) introduced a method to use field of view lines to perform multiple-camera tracking.
The method creates inter-camera meta-target associations when objects enter at the scene edges. They also
said that a plane-induced homography could be used for tracking, but this method was not well described.
Their homography-based system would not work if targets use only one side of a camera to enter the scene.
This paper overcomes this limitation and fully describes a practical homography-based tracker.
A new method to find the feet feature is introduced. The method works especially well if the camera is
tilted, when using the bottom centre of the target's bounding-box would produce inaccurate results. The new
method is more accurate than the bounding-box method even when the camera is not tilted. Next, a method
is presented that uses a series of corresponding point pairs "dropped" by oblivious, live human targets to find
a plane-induced homography. The point pairs are created by tracking the feet locations of moving targets that
were associated using the field of view line method. Finally, a homography-based multiple-camera tracking
algorithm is introduced. Rules governing when to create the homography are specified. The algorithm ensures
that homography-based tracking only starts after a non-degenerate homography is found. The method works
when not all four field of view lines are discoverable; only one line needs to be found to use the algorithm. To
initialize the system, the operator must specify pairs of overlapping cameras. Aside from that, the algorithm
is fully automatic and uses the natural movement of live targets for training. No calibration is required.
Testing shows that the algorithm performs very well in real-world sequences. The consistent labelling
problem is solved, even for targets that appear via in-scene entrances. Full occlusions are handled. Although
implemented in Matlab, the multiple-camera tracking system runs at eight frames per second. A faster implementation
would be suitable for real-world use at typical video frame rates.