Stereo Vision for Autonomous Underwater Robotic Systems

Abstract

Stereo cameras are becoming increasingly important for underwater autonomous robotic systems. These are a low cost and small-sized solution for capturing dense geometric information about the robot’s environment. While algorithms to calculate stereo depth in air are well studied and not significantly impacted by disturbances in the robot's environment, getting accurate and generalized underwater stereo depth for navigation (especially collision avoidance) remains a challenge. Matching algorithms used for stereo depth are extremely sensitive to distortions, lack of clarity, extreme changes in lighting - all of which are features of the underwater environment. In this thesis, we work toward addressing this challenge. First, we create a large data set which captures underwater and in-air images at the same spot in a tank. Through this data set, we can study numerous mappings between images of the same scene in air and underwater. We also vary lighting conditions as we collect this data. Second, we use this data to train a conditional generative adversarial network (pix2pix) that can generate in-air equivalent of a given underwater image. Finally, we use the trained pix2pix model to convert raw underwater stereo pairs to their generated undistorted in-air equivalents, which can then be used to create disparity maps and depth projections of the underwater scene using in-air camera calibration parameters. Using these disparity maps and depth projections, we are able to obtain trajectories that successfully avoid underwater obstacles.