A Sonar-based AUV Positioning System for Underwater Environments with Low Infrastructure Density

TL;DR

This work tackles robust localization for Autonomous Underwater Vehicles in sparsely instrumented underwater environments by introducing a real-time sonar-based global positioning pipeline that fuses INS data with Forward-Looking Sonar observations inside a Particle Filter operating over $SE(2)$. It leverages two complementary frontends—a YOLO-based sonar asset detector (SAD) and a YOLOv8 place-recognition module (PRec)—to resolve observation ambiguities, including symmetric asset configurations. The system localizes using known asset poses and simplified CAD models to estimate $[x, y, \theta]$, and fuses observations to improve convergence and accuracy, as demonstrated in Gazebo/DAVE virtual experiments. Early results indicate that combining SAD and PRec yields superior localization robustness over using SAD alone, motivating future real-world validations and real-to-sim transfer development to close the sim-to-real gap.

Abstract

The increasing demand for underwater vehicles highlights the necessity for robust localization solutions in inspection missions. In this work, we present a novel real-time sonar-based underwater global positioning algorithm for AUVs (Autonomous Underwater Vehicles) designed for environments with a sparse distribution of human-made assets. Our approach exploits two synergistic data interpretation frontends applied to the same stream of sonar data acquired by a multibeam Forward-Looking Sonar (FSD). These observations are fused within a Particle Filter (PF) either to weigh more particles that belong to high-likelihood regions or to solve symmetric ambiguities. Preliminary experiments carried out on a simulated environment resembling a real underwater plant provided promising results. This work represents a starting point towards future developments of the method and consequent exhaustive evaluations also in real-world scenarios.

Figures (9)

• Figure 1: The top image shows a typical trajectory (poses 1 to 8) followed by an AUV while inspecting a human-made asset, which was imitated during the virtual experiments. The bottom image shows a graphical representation of a map of an underwater plant. Approximate 3D CAD models of each "object" (i.e., asset) included in the map are also provided.
• Figure 2: Example of the sonar's projection model. Here it is evident that all the points that live in the arc $\phi$ will be projected in the same point.
• Figure 3: Example of the YoloOBB detection pipeline using a sonar image produced utilizing the simulator Dave zhang2022dave. The OBBs produce a more informative measurement with respect to the classic BBs.
• Figure 4: Representation of the 2D grid centered on the asset used to produce the dataset for training the two different networks. The green dots represent valid poses in which there is no collision, while the red dots represent poses to be avoided due to collisions.
• Figure 5: On the left it is showcased a situation in which there is a symmetric underwater asset, and the two poses from which the sonar images computed will result to be almost the same. On the right, it is shown sonar images produced by the simulator in which the results of the detection are ambiguous.