Robust LiDAR-Stereo Fusion for Lunar Autonomy

The Artemis mission requires sensing technologies and associated algorithms to enable vehicles to operate autonomously with fast cadence and high reliability. In support of fast cadence and high reliability autonomy, Quartus Engineering will develop a LIDAR-Stereo fusion system using state of the art algorithms optimized for lunar worthy computing platforms. Fusion of LiDAR and stereo image data using state of the art algorithms can enable accurate long range obstacle detection and mapping while preserving dense visual and texture information. Quartus will initially prototype and compare three fusion methods: LiDAR-guided stereo matching, depth completion and fusion techniques, and deep learning-based fusion. Quartus will leverage the NASA POLAR Stereo dataset and LunarSim to generate synthetic LiDAR-Stereo data that closely mimics lunar conditions. This simulated data will be used to evaluate different algorithmic approaches and downselect a particular fusion and stereo matching pipeline. The downselected fusion and stereo matching pipeline will be re-implemented as an optimized C++ library, with highly parallelizable portions of the algorithms factored into subroutines with support for RISC-V Vector, GPU, or FPGA accelerators. This support for hardware accelerators will be achieved through a combination of OpenCL and Microchip’s SmartHLS compiler software. Finally, Quartus will select an off the shelf RISC-V computer with comparable properties to anticipated lunar worthy RISC-V platforms and evaluate algorithm performance on this test hardware to demonstrate that desired robustness, accuracy, and latency can be achieved on a representative computing device. This work is the first step towards commercializing a LiDAR-Stereo module comprised of a LiDAR system, a stereo system, precise calibrations, and a lunar worthy compute device hosting robust, high accuracy, low latency algorithms for stereo matching and LiDAR-stereo fusion.