Automatic unconstrained online configuration of a master-slave camera system

Master-slave camera systems - consisting of a wide-angle master camera and an actively controllable pan-tilt-zoom camera - provide a large field of view, allowing monitoring the full situational context, as well as a narrow field of view, to capture sufficient details. Unconstrained calibration of such a system is a non-trivial task. In this paper a fully automatic and adaptive configuration method is proposed. It learns a motor map relating image coordinates from the master view to motor commands of the slave camera. First, a rough initial configuration is estimated by registering images from the slave camera onto the master view. In order to be operational in poorly textured environments, like hallways, the motor map is online refined by utilizing correspondences originating from moving objects. The accuracy is evaluated in different environments, as well as in the visual and the infrared spectrum. The overall accuracy is significantly improved by the online refinement