DOC 04PERCEPTION

Imaging sonar and acoustics

LEVIATHAN's canonical forward-looking imaging sonar produces a range × azimuth image of relative acoustic return. Every cell records which surfaces contributed to it and how valid its value is. It has a CPU scientific reference and a GPU backend that reproduces it bit for bit.

Two sonar paths

Beam geometry

All geometry comes from the sensor snapshot. Bearings span the azimuth field of view and elevations span the vertical aperture, both on inclusive grids centred on the sensor's +x boresight. A mount tilt lives only in the sensor's mount transform. Each (bearing, elevation) ray is cast from the sensor origin against the frozen canonical scene. The nearest surface wins, so nearer geometry shadows what lies behind it along that ray.

Return model

Each ray is one quadrature sample of solid angle dΩ = cos(e) Δb w_e. The trapezoid weights w_e of the elevation grid sum exactly to the aperture. A ray that hits a surface at slant range R and incidence θ contributes

G

one-way intensity beam pattern, uniform or Gaussian with stated −3 dB full widths; two-way is G²

σ_b(θ)

the surface's acoustic response; the implemented model is Lambert, μ cos²θ

R² dΩ / cos θ

the area of surface the ray sample represents

α

absorption by Francois and Garrison (1982) at the hardware frequency, for the declared water temperature, salinity and pH, at the sensor's depth

1/R⁴

two-way spherical spreading in intensity

Every ray whose range falls in a cell adds its contribution to that cell, so returns are summed incoherently over the elevation aperture. The sum is the cell's linear_return. An optional fully developed speckle multiplies it by an exponential random variable with unit mean, drawn from keyed randomness, and is reported as a separate channel.

Per-cell ground truth

A sonar cell can contain contributions from several surfaces, at different elevations and ranges within the bin. LEVIATHAN keeps that information.

• Primary contributor. The dominant sample in the cell, with its instance, object, triangle, exact slant range, elevation, incidence, world hit point, normal, semantic label, acoustic status and whether it belongs to a moving body.
• Top-K contributors. Up to K contributors per cell in a deterministic order: dominance, then range, elevation, instance and triangle. Each carries its contribution, geometric weight and energy fraction.
• Counts and truncation. The number of samples integrated into the cell, and the share of its return that falls outside the top-K list.

Dominance is the contribution itself when the cell's radiometry is fully known. When radiometry is unknown, it is the geometric weight, which is everything except the surface response. When the medium is unsupported, it is the geometric weight without absorption. The primary_basis channel records which rule was used. Only samples the model actually integrated are contributors.

Validity

Acoustic materials

Acoustic materials are kept separate from optical reflectance and are never derived from it. Each parameter carries its meaning, units, source and evidence class. The library contains only clearly labelled test and reference surfaces, such as Lambert surfaces with stated μ values. No published high-frequency backscatter parameter for seabed or structure materials could be verified from a primary source, so no physical material is claimed.

In practice, real scene content without a declared acoustic material is reported as acoustic property unknown. Its geometry ground truth, including range, point, normal and identity, is still valid and usable, but its return is not given a value.

Configuration

A canonical sonar is declared on a vehicle in the canonical vehicle catalogue, with its hardware and sampling schedule. In the current release, the built-in production vehicles declare a front camera only. The catalogue's reference survey vehicle is explicitly labelled as a test and reference model. It declares a 900 kHz imaging sonar with a 90° azimuth field of view, a 20° elevation aperture and a Gaussian beam, sampled at 5 Hz, alongside a DVL and an IMU. Its rates and hardware are declared test values, not a device specification.

CPU reference and GPU backend

The CPU implementation is the scientific authority. Before any GPU code existed, its output was frozen for 28 canonical scenes. The fixture covers flat and tilted surfaces, known, unknown and unsupported materials, an unsupported medium, static and moving targets, three kinds of occlusion, shadowing, an exact coincident-surface tie, K = 1, 2, 4 and 16, near and far range windows, the sea-surface interface, 1 to 8 elevation samples, and speckle. It has never been regenerated.

The GPU backend shares the CPU path's preparation and output stages. It accelerates instance culling and bounding-volume traversal on the device in float64, with the same operation order, and evaluates the per-sample physics on the host with the reference's own expressions.

Single-precision traversal was evaluated and rejected. It is not bit-identical, so it would break exact dominance ties and make triangle selection at edges depend on the platform.

Analytic sonar tools

For interactive work, the Dashboard's Imaging sonar panel renders an analytic range × bearing image from a pose, as a B-scan or a fan. It reports the image grid, range window, field of view, dynamic range, empty and saturated cells, a measured range-falloff fit, range and cross-range resolution, sound speed, absorption at the operating frequency and the maximum ping rate.

The panel is candid about what an image can support. In the capture above it reports that the falloff fit explains only 17 % of the variance, because an image of discrete targets in open water has no reverberation floor to measure.