RGB-LiDAR Pipeline for 3D Bounding Box Estimation in Low SWaP-C Indoor Navigation Applications

Abstract

The generation of 3D bounding boxes from a combination of RGB images and depth data is an important area of research for autonomous systems. Additional constraints must be considered in order for any method to be implemented in a small form factor unmanned aerial vehicle (UAV) or other robotic system. In this work, an "ultra-lightweight" pipeline is developed and used to generate 3D bounding boxes from aligned RGB-LiDAR images. Different from current implementations, the design of this pipeline was made to maintain a similar performance to more computationally intensive algorithms while also being implementable onboard low SWaP-C systems. The pipeline is demonstrated in a computationally restricted indoor environment by a small rover to navigate around obstacles while searching for a target object.