Deep Hough Voting based 3D object detection and pose estimation in LiDAR point clouds

Spatial reasoning, particularly the ability to detect and recognize objects within a 3D
environment, is crucial for robotic systems aiming to navigate and function in unfamiliar settings. In this thesis, we introduce a method specifically designed for detecting cuboid-shaped objects within point clouds through a voting-based mechanism and estimating their 9DoF pose. Our detection framework is accompanied by a synthetic data generation pipeline, which is utilized to generate the necessary training data. Our evaluation reveal that our method exhibits robust performance when applied to real- world data, even though it was exclusively trained on synthetic data. The examination of the sim-to-real gap shows just minimal degradation of orientation estimation and a moderate decline in detection capability. We test different orientation representations and propose a way to map equivalent but distinct orientations of cuboids to a single canonical orientation in a deterministic way.