Occlusion Modeling for Coherent Echo Data Simulation: A Comparison Between Ray-Tracing and Convex-Hull Occlusion Methods

The ability to simulate realistic coherent data sets for synthetic aperture imaging systems is crucial for the design, development, and evaluation of sensors and their signal processing pipelines, machine learning algorithms, and autonomy systems. In the case of synthetic aperture sonar (SAS), collecting experimental data is expensive, and it is rarely possible to obtain ground truth of the sensor's path, the speed of sound in the medium, and the geometry of the imaged scene. Simulating sonar echo data allows signal processing algorithms to be tested with known ground truth, enabling rapid and inexpensive development and evaluation of signal processing algorithms. The de facto standard for simulating conventional high-frequency (i.e., $> {\text{100}}$ kHz) SAS echo data from an arbitrary sensor, path, and scene is to use a point- or facet-based diffraction model. A crucial part of this process is acoustic occlusion modeling. This article describes a SAS simulation pipeline and compares implementations of two occlusion methods: 1) a ray-tracing method and 2) a newer approximate method based on finding the convex hull of a transformed point cloud. The full capability of the simulation pipeline is demonstrated using an example scene based on a high-resolution 3-D model of the SS Thistlegorm shipwreck, which was obtained using photogrammetry. The 3-D model spans a volume of $\text{220}\times \text{130}\times \text{25}\,\text{ m}$ and is comprised of over 30 million facets that are decomposed into a cloud of almost 1 billion points. The convex-hull occlusion model was found to result in simulated SAS imagery that is qualitatively indistinguishable from the ray-tracing approach and quantitatively very similar, demonstrating that the use of this alternative method has potential to improve speed while retaining high fidelity of simulation. The convex-hull approach was found to be up to four times faster in a fair speed comparison with serial and parallel central processing unit (CPU) implementations for both the methods, with the largest performance increase for wide-beam systems. The fastest occlusion modeling algorithm was found to be graphics processing unit (GPU)-accelerated ray tracing over the majority of scene scales tested, which was found to be up to two times faster than the parallel CPU convex-hull implementation. Although GPU implementations of convex-hull algorithms are not currently readily available, the future development of GPU-accelerated convex-hull finding could make the new approach much more viable. However, in the meantime, ray tracing is still preferable, since it has higher accuracy and can leverage the existing implementations for high-performance computing architectures for better performance.