A Robust Uncalibrated Visual Compass Algorithm from Paracatadioptric Line Images

Due to their wide panoramic field of view, paracatadioptric cameras are becoming ubiquitous in many robotic applications. A challenging problem consists in using these vision sensors as a visual compass, that is to exploit the image data solely to provide an estimate of their rotational motion when mounted on a mobile/humanoid robot. Existing visual compass algorithms are difficult to implement in real scenarios since they assume known camera calibration parameters and known geometry of image features (e.g. correspondence between points or parallelism among between lines). In this paper we present a new multi-view property for paracatadioptric cameras called disparity-circles constraint and we use it to design a visual compass algorithm to estimate the z−axis camera rotation angle. The proposed algorithm only uses the image projection of 3-D parallel lines and is suitable for real-time implementation. Those 3-D lines that are parallel are shown to be automatically detectable via a RANSAC implementation of the algorithm. Real-data experiments conducted with a paracatadioptric camera mounted on a mobile robotic platform prove the robustness of the proposed approach