LiDAR Point Cloud Object Detection

An autonomous vehicle's roof-mounted spinning LiDAR sweeps 360° around it dozens of times a second, each rotation producing tens of thousands of unlabeled 3D returns. Before any object can be tracked, the perception stack has to answer two questions with no ground-truth to lean on: which points are the road, and which points belong together as one obstacle? This simulator runs the real answer to both — a RANSAC plane fit that discovers the ground surface by consensus among random point triples, and a voxel-grid connected-component pass that clusters the leftover points into discrete objects — live, on a synthetic scene of cars and pedestrians you can reshuffle and a sensor you can reconfigure.