Distributed 3-D localization in camera networks

Consider a problem in which N cameras in a camera sensor network must estimate their pose in a distributed fashion exclusively from image measurements. Using standard computer vision techniques, each camera can extract feature points from the images, match these points across a few neighboring cameras, and estimate the relative rotation and translation (up to a scale factor) of the neighboring cameras. In this paper, we propose distributed consensus algorithms in SE(3) for estimating a globally consistent pose for all N cameras. Traditional consensus algorithms have been developed under the assumptions that the measurements are Euclidean and low-dimensional, and consensus is performed directly on the measurements. In our problem, the measurements (images) are high-dimensional and consensus must be performed on a quantities (camera poses) that are non-Euclidean are different from the measurements. This poses several challenges to traditional consensus for computer vision applications, which we address in this paper. While our methods are designed specifically with the camera sensor network application in mind, our results are applicable also to localization problems in a more general setting. We provide synthetic simulations to test the validity of our approach.