Deviations Detection in Registered 3D Point Clouds

This paper explores the validity of three distance metrics: a Point-to-Point distance, a Point-to-Plane distance and a Modified Hausdorff Distance, in detecting the misplacement of a Time-of-Flight camera in a registered 360-degree surround view perception system installed on a vehicle. The vehicle’s undamped mechanical vibrations can unexpectedly pivot one of the cameras to an unforeseen orientation, leading to inaccurate processes that rely on correct registered point clouds. Parameterizing the camera exterior 3D orientation and synthesizing point clouds of a city model inside Blender shows that a pair of registered point clouds has a unique distance value in the camera configuration space for each metric method. This allows continuous monitoring over time and notifies the user of any sudden changes in the camera configuration.