Bathymetric Surveying Through Neural Inverse Sonar Modeling

QC 20240516

Abstract

Autonomous underwater vehicles (AUVs) equipped with sidescan sonars (SSS) and remotely operated vehicles (ROVs) equipped with forward looking sonars (FLS)  have become vital tools for many underwater applications, among which, bathymetric mapping is one of the most important yet challenging tasks. Providing high-resolution imagery, SSS and FLS are particularly suitable for compact, low-cost and scalable vehicles, however their linear array design makes their measurements ambiguous in elevation. This limitation, makes inverse sonar modeling necessary for accurate and detailed bathymetric surveying in open sea missions. Solving such an inverse problem is usually ill-posed, even with repeated observations from different distances and viewpoints. Thus, this thesis has focused on learning-based approaches to sonar modeling, aiming to leverage advances in recent deep learning.

We present our contributions in three areas tackling this inverse sonar modeling. First, our work on learning SSS model with data-driven approaches shows how convolutional neural networks (CNNs) can be successfully applied to learn the inverse model in an end-to-end, supervised-learning fashion. We further show how to fuse different CNN estimates depending on the representations used for the seabed. We demonstrate that with an explicit grid representation, the uncertainty estimates of CNN's predictions can be of help in the fusion, while an implicit neural representation, specifically, a neural heightmap parameterized by multi-layer perceptrons (MLPs) could handle the fusion implicitly by posing the problem as a global optimization.Secondly, we leverage the methodology of representing the seabed with implicit neural representations and propose to use a Lambertian model based on surface rendering for sonar modeling. The proposed approach does not require collecting ground truth bathymetry and can be used on different datasets with different sensor setups. The experiments also demonstrate how the approach can converge  to a self-consistent map without any external bathymetric data.Finally, we adapt two differentiable volume rendering techniques  in  computer graphics to sonar modeling and show their advantages over surface rendering  on accurately modeling the physics behind the sonar ensonification process. In specific, a soft rasterization-based renderer with explicit mesh representations, and a ray-casting-based volumetric rendering with implicit neural representations with parametric encodings. The latter solution, in particular, is capable of leveraging deep learning advances without Lambertian approximation.  Our results show the proposed approach not only outperforms surface rendering solutions but its parametric encoding also allows it to outperform volume rendering methods with non-parametric encodings. We also demonstrate the potential application of increasing the resolution of  a low-resolution prior map with FLS data from low-altitude surveys.

urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-346474