Abstract
This paper deals with a number of computer vision techniques for the integration and interpretation of visual cues in RGB and forward-look sonar images. To utilize a low-cost GoPro stereo imaging system with dedicated water-proof housing and an Oculus sonar, we perform camera calibration by ray tracing to account for the impact of refraction at the housing glass ports, and calculate the relative poses of all three imaging systems. Utilizing the data for our calibrated system, we describe and assess certain 3-D reconstruction methods to determine the relative position of various scene targets for collision avoidance, to generate 3-D object models, and to enhance RGB images by haze removal. Experiments with the real images of various targets in a pool and a water tank under both good visibility and turbidity are presented to demonstrate some advantages in the integration of multi-modal visual cues. Collectively, these methods are targeted for the realization of capabilities that enhance marine robotics perception and autonomy in near-seabed operations.