Terrain constrained stereo correspondence

Abstract

There is a persistent need in the oce¿anographie community for accurate three dimensional reconstructions of seafloor structures. To meet this need underwater mapping techniques have expanded to include the use of stereo vision and high frequency multibeam sonar for mapping scenes 10's to 100's of square meters in size. Both techniques have relative advantages and disadvantages that depend on the task at hand and the desired accuracy. In this paper, we develop a method to constrain the often problematic stereo correspondence search to small sections of the image that correspond to estimated ranges along the epipolar lines calculated from coregistered multibeam sonar micro-bathymetry. This approach can be applied to both sparse feature based and dense area-based stereo correspondence techniques. Data were collected on an underwater vehicle survey using a calibrated stereo rig and a multibeam sonar gathering coincident datasets. Overall, the constrained correspondence method shows improvements in the number and reliability of correct matches and allows for reduction in complexity of feature descriptors but it is heavily reliant on the quality of the intrinsic and extrinsic calibration of the camera and sonar systems.