HydroMap: Probabilistic Water Surface Elevation Mapping for Semantic Scene Representation in Inland Waterways
HydroMap is an odometry-decoupled framework that fuses stereo observations with uncertainty propagation to generate a persistent, probabilistic 2.5D elevation map of water surfaces, effectively complementing sparse LiDAR data for autonomous navigation in inland waterways.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a boat moving through a narrow canal, surrounded by old stone walls, bridges, and overhanging trees. For a human captain, the task is intuitive: look at the water to know where the boat floats, and look up to see how much space exists under a bridge. But for a computer trying to steer the same boat, the water is a blind spot. Modern autonomous vessels rely on lasers that bounce off solid objects to build a 3D map of their world. These lasers work beautifully on stone and steel, but when they hit the water, the light often vanishes or reflects in confusing ways. The result is a digital map that shows the walls and the bridge perfectly, but leaves a gaping, invisible hole where the water surface should be. Without knowing exactly where the water is, the computer cannot tell if the boat is floating safely or if it will crash into a low-hanging branch.
To solve this, researchers have developed a new system called HydroMap, which acts as a persistent memory for the water's surface. Instead of relying on lasers to see the water, the system uses standard stereo cameras—two lenses spaced apart like human eyes—to measure the height of the water in real time. The key innovation is that this system does not just guess the water level for a single moment; it builds a lasting, probabilistic map that remembers the water's shape over time, even as the boat moves and the view changes. By combining these camera observations with the laser map of the surrounding structures, the system creates a complete picture of the world, filling in the invisible gap with a reliable estimate of the water's elevation.
The researchers tested this approach on two different waterways: the Pohang Canal in South Korea and the Leuven Vaart in Belgium. In these environments, the water conditions varied from calm to slightly choppy, and the clarity of the water ranged from clear to quite murky. In the murky conditions of the Belgian canal, the laser scanners failed to pick up any returns from the water surface at all, leaving the map completely empty in those areas. The HydroMap system, however, successfully reconstructed the water surface using only the camera images. When the researchers compared the computer's map against the few laser returns that were available for verification, the height estimates were remarkably accurate. The difference between the estimated water level and the actual laser measurement was less than 5 centimeters on average, with the vast majority of errors staying well under 10 centimeters.
A critical part of this success is how the system handles uncertainty. Because camera measurements can be imperfect, and because the boat's own position is never known with absolute precision, the system does not simply average the numbers. Instead, it tracks two types of uncertainty for every point on the water: the error specific to that single camera snapshot, and the shared error that comes from the boat's movement. By keeping these distinct, the system can fuse thousands of observations together without becoming overconfident. It knows when to trust a new measurement and when to rely on what it has already learned, creating a map that updates twice per second for the water height and once per second for the full scene.
The final output is a unified 2.5D map that shows not just the water, but also the boundaries where the water meets the land and the space above the boat. This allows the autonomous vessel to answer vital questions instantly: Is the water deep enough here? Is the bridge high enough to pass under? In one specific test near a bridge in Belgium, the system estimated the clearance to be 6.90 meters, which matched a manual human measurement of 6.85 meters almost exactly. This level of precision means the boat can navigate narrow channels and pass under obstacles with the same spatial awareness a human captain would have, but with the added benefit of a constantly updated digital record of the environment.
While the system is highly effective, the researchers are careful to note its limits. The camera-based approach can struggle if the water is perfectly smooth and reflective, or if the lighting is poor, which might cause the system to miss parts of the surface or confuse reflections for solid objects. Additionally, the current model assumes the water is relatively still, so it does not yet account for the complex, shifting patterns of waves created by the boat's own wake or strong currents. Despite these constraints, the work demonstrates that by combining the strengths of different sensors—using lasers for the solid world and cameras for the liquid one—it is possible to create a complete, persistent map of an inland waterway. This capability is a significant step forward for autonomous navigation, turning a previously invisible hazard into a clearly defined part of the map.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.