Unsupervised camera network structure estimation based on activity

In this paper we consider the problem of unsupervised topology reconstruction in uncalibrated visual sensor networks. We assume that a number of video cameras observe a common scene from arbitrary and unknown locations, orientations and zoom levels, and show that the extrinsic and calibration matrices, fundamental and essential matrices, the homography matrix, and the physical configuration of the cameras with respect to each other can be estimated in an unsupervised manner. Our method relies on the similarity of activity patterns observed at various locations, and an unsupervised matching method based on these activity patterns. The proposed method works in cases with cameras having significantly different orientations and zoom levels, where many of the existing methods cannot be applied. We explain how to extend the method to a multicamera case where more than two cameras are involved. We present both qualitative and quantitative results of our estimates, and conclude that this method can be applied in wide area surveillance applications in which the deployed systems need to be flexible and scalable, and where calibration can be a major challenge.