Stereo rig with wide baseline is necessary when accurate depth estimation for distant object is desired. However, in order to make calibration pattern to be viewed from both left and right cameras, the wider the baseline the bigger the calibration pattern is required. In contrast to the traditional stereo calibration method using calibration pattern, we propose a self-calibration approach that can estimate cameras' rotation matrices for stereo rig with wide baseline (3 m). Given images taken from left and right cameras, the relative roll and pitch angles between two cameras are recovered by aligning sea horizon in left and right images. The pitch angle is estimated by making the projections of one point at infinite distance appear at the same location in both images. A photometric minimization is applied to refine the rotation parameters. Compared with conventional checkerboard-based calibration techniques which require extra equipments or personnel, our approach only needs a pair of sea images. Moreover, unlike most self-calibration approaches, feature detection and matching are not required which makes it possible to apply our approach on featureless images. As a result, it is flexible and easy to implement our approach on sea surface images. Real world experiments demonstrate the feasibility of our approach.
Oceanic Eddy Identification Using an AI Scheme