Motion Estimation-based Self-Calibration of a Log-polar Pan-Tilt Camera

Two topics intimately linked to active vision are spacevariant vision and mechanical devices. On the one hand, log-polar images is a biologically inspired space-variant imaging model featuring very attractive properties. On the other hand, pan-tilt heads are a common platform in active vision, allowing the observer to dynamically perceive the environment. In order to relate motor commands (for the pan-tilt mechanism) to displacements in the image plane, such a mapping have to be worked out somehow. Often, a calibration process is performed by using artifacts such as calibration patterns. On the contrary, we follow the principles of active vision and propose a technique for the selfcalibration of the camera. The only requirement is a scene with some textured parts. The process is based on a motion estimation algorithm which operates directly on log-polar images, which makes the approach feasible to be used in real log-polar sensors mounted on some robotic set-up.