Slide 13 of 27
Once we can track moving objects in each camera, we can use that information to automatically calibrate all the cameras. Using computer vision methods, we can match the trajectories of the moving objects, and use that correspondence to determine how the cameras are positioned. This can be adapted over time to keep the cameras fully calibrated.
For example, here are two views of a scene, taken 180 degrees apart, using different cameras. Also shown are the trajectories of the moving objects detected in the two views. By matching these trajectories, we recover the relative positions of the cameras. To show how well this works, we have taken the image from one camera, and automatically warped it to reflect the view that should have been seen from the other camera. One can see that the alignment is very good.