Parameter-unaware autocalibration for occupancy Mapping

People localization and occupancy mapping are common and important tasks for multi-camera systems. In this paper, we present a novel approach to overcome the hurdle of manual extrinsic calibration of the multi-camera system. Our approach is completely parameter unaware, meaning that the user does not need to know the focal length, position or viewing angle in advance, nor will these values be calibrated as such. The only requirement to the multi-camera setup is that the views overlap substantially and are mounted at approximately the same height, requirements that are satisfied in most typical multi-camera configurations. The proposed method uses the observed height of an object or person moving through the space to estimate the distance to the object or person. Using this distance to backproject the lowest point of each detected object, we obtain a rotated and anisotropically scaled view of the ground plane for each camera. An algorithm is presented to estimate the anisotropic scaling parameters and rotation for each camera, after which ground plane positions can be computed up to an isotropic scale factor. Lens distortion is not taken into account. The method is tested in simulation yielding average accuracies within 5cm, and in a real multi-camera environment with an accuracy within 15cm.