A Hardware Accelerated Computer Vision Library for 3D Reconstruction Onboard Small Satellites

Here we present benchmarks and the expected performance of the SSRLCV (Small Satellite Research Laboratory Computer Vision) library that will be tested in Low Earth Orbit onboard the 6U MOCI (Multiview Onboard Computer Vision) cube satellite. The SSRLCV software library, used on the MOCI cubesat, is written in CUDA and C++ for Nvidia GPU/SoCs and performs structure from motion to generate 3D terrain information from a series of locally generated orbital images. Scale and Rotation invariant features are extracted from images, matched between those images, then an initial 3D point cloud is estimated with feature triangulation. Noise is removed from the point cloud and a gradient descent method, known as bundle adjustment, is used to refine the estimated camera parameters and location information of the satellite. The research simulates satellite imagery from LEO with 3D rendering software to test image data. Tests are run on the Nvidia TX2 and TX2i with timing, state, and power usage tracking. Reconstruction accuracy is measured by volumetric comparison and an Iterative Closest Point algorithm to allow for comparison to ground truth 3D models. The results show accurate 3D reconstruction of the surface of Earth feasible within 15 to 100 meters, depending on the camera system and altitude, while maintaining favorable power usage and computation time.