Exploiting natural contours for automatic sonar-to-video calibration

Abstract

The paper addresses the problem of coregistration of data acquired by two distinct ego-centric exteroceptive sensors mounted in an autonomous underwater vehicle: a video camera and a mechanically scanning profiler sonar. As it is the case for the http://www.laas.fr/reforme-CNRS.html experimental platform used in this study1, these two perception sensors can often be installed in flexible geometric configurations on the robot's crash frame, and thus their relative orientation and displacement may change from mission to mission, even if during each mission they are kept fixed. The algorithm exploits the presence in the sea bottom of a boundary between two regions that can be discriminated both in the video frames and on the sonar profiles, and requires no specially tuned artificial settings. The registration is solved in a completely autonomous manner, using the ability of the robot to track the natural contour based on the video data. Our algorithm finds the mapping from the sonar coordinates to the image coordinates that minimizes the average entropy of a probabilitic model that is fit to the inverse mapping of the segmented images. It does not require reconstruction of the robot's trajectory during data acquisition, and is thus insensitive to positioning problems.