Large-Scale Dense 3D Mapping Using Submaps Derived From Orthogonal Imaging Sonars

TL;DR

This work advances underwater 3D mapping by introducing submapping with orthogonal wide-aperture imaging sonars to produce dense 3D maps without reliance on prior object models. It combines SLAM with fused orthogonal sonar observations, object-specific Bayesian inference, and a novel submap construction that retains data between SLAM keyframes to increase coverage. Through simulations and real-world tests, the authors show that inference-based mapping performs best when repeating simple objects are abundant and a trained semantic model is available, while submapping delivers robust, high-coverage maps in complex environments with limited or no prior knowledge. The approach broadens practical deployment of autonomous underwater mapping in fully unknown scenes and highlights future directions for combining inference with submapping to leverage both complex geometry and object regularities.

Abstract

3D situational awareness is critical for any autonomous system. However, when operating underwater, environmental conditions often dictate the use of acoustic sensors. These acoustic sensors are plagued by high noise and a lack of 3D information in sonar imagery, motivating the use of an orthogonal pair of imaging sonars to recover 3D perceptual data. Thus far, mapping systems in this area only use a subset of the available data at discrete timesteps and rely on object-level prior information in the environment to develop high-coverage 3D maps. Moreover, simple repeating objects must be present to build high-coverage maps. In this work, we propose a submap-based mapping system integrated with a simultaneous localization and mapping (SLAM) system to produce dense, 3D maps of complex unknown environments with varying densities of simple repeating objects. We compare this submapping approach to our previous works in this area, analyzing simple and highly complex environments, such as submerged aircraft. We analyze the tradeoffs between a submapping-based approach and our previous work leveraging simple repeating objects. We show where each method is well-motivated and where they fall short. Importantly, our proposed use of submapping achieves an advance in underwater situational awareness with wide aperture multi-beam imaging sonar, moving toward generalized large-scale dense 3D mapping capability for fully unknown complex environments.

Figures (11)

• Figure 1: Example sonar geometry. Sensor origin is shown at left, with max range denoted. The sensor swath, or horizontal field of view is shown as $\theta$ and the vertical aperture is shown as $\phi$. We note that an example $\theta$ and $\phi$ are shown to point $p$.
• Figure 2: SOCA-CFAR overview. Purple cells show training cells, blue the guard cells, and red the cell under test.
• Figure 3: Mapping via object specific Bayesian inference block diagram. A pair of orthogonal sonar images is provided as input (black lines bound the region of overlap between the two sonar fields-of-view). The images are processed according to Section \ref{['sonar-fuse']}. The horizontal image is segmented as in Section \ref{['sonar-seg']} (colors denote different object classes - seawall in green, rectangular pilings in yellow, cylindrical pilings in red). The resulting 3D points enrich each object's model (Eq. \ref{['eq:bayes rule']}), while MAP inference is applied to 2D points (Eqs. \ref{['eq:query']}, \ref{['eq:query2']}). We then use the planar SLAM solution to register the resulting point cloud. The synthetic sonar images shown here are sampled from the virtual environment depicted in Fig. \ref{['fig:marina_2_pic']}.
• Figure 4: Simulation Environments.
• Figure 5: Simulation coverage results. Coverage in both table and color format. Each cell reports voxel count, with color mapping from blue to orange as low to high coverage. The vertical axis shows the varying keyframe rotations in degrees, with the horizontal axis showing the Euclidean distance between keyframes in meters. Each system type is shown here, with sonar fusion mapping at left, inference based mapping at center and submapping at right. Subfigure (a) shows the results for simulated marina 1, (b) shows the results for simulated marina 2, (c) shows the results for the simulated harbor and (d) shows the results for the simulated plane. Note, the voxel count is colored according to the scale bar given at the right of each subfigure.