Active Mapping of Underwater Litter Using Camera-Sonar Fusion

Can an uncrewed underwater vehicle actively combine camera and sonar observations to find static seafloor litter faster than predefined coverage or single-sensor search?

Existing underwater litter surveys either follow fixed paths such as lawnmower patterns and waste time re-observing confirmed areas, or rely on a single sensor: cameras degrade in turbidity and low light, while sonar is noisy and low-resolution. This makes small, visually ambiguous debris hard to locate efficiently.

Prior approaches use predefined coverage paths, single-sensor acoustic or visual detection, or information-driven planners that reduce overall map uncertainty. They do not fuse camera and sonar for active seafloor target search, and generic entropy reduction does not prioritize confirming likely object candidates.

The method keeps an Octomap Bayesian occupancy grid and updates it with Mask R-CNN detections from a camera and a forward-looking sonar. Each sensor is represented by learned range- and bearing-dependent true-positive and false-positive probability tables. At each step, the robot enumerates candidate waypoints (6 m grid, 8 headings; 3 m grid for camera-only) and scores them by the sum over visible voxels of entropy plus a 5× bonus for voxels with occupancy probability between 0.7 and 0.95. The UUV then navigates at 0.7 m/s while sampling at 15 Hz camera and 20 Hz sonar.

In a MARUS simulated scene of about 33 m × 24 m with 11 objects in three clusters, dual-sensor active search found all 11 objects after 27 m of travel; the lawnmower found only 4 objects by that distance and required 110 m to find all. Active mapping also found the third cluster without passing directly over it. In single-sensor comparisons, dual-sensor active search was roughly twice as fast as sonar-only or camera-only; the single-sensor variants showed no clear winner. A second scenario with top-aligned clusters preserved the active-search advantage.

The evaluation is simulation-only in MARUS; no real underwater field trials are reported. It assumes known seafloor geometry and known robot pose, treats camera and sonar as conditionally independent, uses a myopic one-step planner, and learns sensor error models from synthetic data with artificially degraded detector confidence. The experiments cover a small number of scenarios and do not report repeated trials. Future work includes joint sensor modeling, non-myopic planning, and field validation.