Bathymetric Reconstruction From Sidescan Sonar With Deep Neural Networks

Abstract

He is currently a Researcher with the Swedish Maritime Robotics project at KTH. His research interests include robotic sensing and mapping, with a focus on probabilistic reasoning and inference. Most of his recent work has been on applications of specialized neural networks to underwater sonar data. In addition, he is interested in system integration for robust and long-term robotic deployments.John Folkesson (Senior Member, IEEE) received the B.A. degree in physics from Queens College, City University of New York, New York, NY, USA, in 1983, and the M.Sc. degree in computer science and the Ph.D. degree in robotics from the Royal Institute of Technology (KTH),

Towards Differentiable Rendering for Sidescan Sonar Imagery

“…They showed that convergence can be improved by gradually increasing the resolution of the predicted bathymetry. Recently, data-driven approaches [5], [6] using CNNs have been proposed to learn the missing elevation directly from SSS images in a supervised-learning fashion. However, the "ground truth" bathymetry is needed for creating a training set, which is not always practical underwater.…”

“…Similarly, for the shadows in the data, the gradient descent procedure would find a intersection between the arc and the seafloor, but the found intersection is occluded. Inspired by [9], [10], we propose to sample a few points along the ray backwards, starting from the intersection and compute the accumulated transmittance at the intersection T (p(ϕ * )) = exp (− r s 0 ρ(u)du), which is used to indicate if p(ϕ * ) is occluded. Here, the particle density ρ is computed as in [9], [10].…”

“…Inspired by [9], [10], we propose to sample a few points along the ray backwards, starting from the intersection and compute the accumulated transmittance at the intersection T (p(ϕ * )) = exp (− r s 0 ρ(u)du), which is used to indicate if p(ϕ * ) is occluded. Here, the particle density ρ is computed as in [9], [10]. We use T (p(ϕ * )) to weight the predicted SSS intensity, since if T (p(ϕ * )) is close to zero, it indicates that the intersection point is occluded, which corresponds to shadows in the sidescan images.…”

“…As a result, many approaches from computer vision and computer graphics for 3D reconstruction using camera images can be adapted for the same task using sonars. Examples of these are, in the early days, shapefrom-shading (SfS) techniques [1]- [4], and more recently data-driven methods using convolutional neural networks (CNNs) [5], [6] and inverse rendering based on implicit neural representations [7]- [10].…”

Testing the Effect of Bathymetric Data Reduction on the Shape of the Digital Bottom Model

“…A complete study of our planet’s topography is an extremely challenging task, since almost three-quarters of the planet’s surface is covered by water, and the depth in some areas can reach several kilometers [ 6 ]. At present, modeling of the seafloor surface can be performed based on measuring the depth of a body of water, resulting in a corresponding bathymetric dataset [ 7 , 8 , 9 , 10 ].…”