PPC-US-LSF3DOD: A Pseudo-Point-Clouds based Unsupervised and Scalable Late-Sensor-Fusion Method for 3D Obstacle Detection*

To get the 3D coordinates and sizes of multi-scale obstacles, we propose an unsupervised and scalable late sensor fusion method using pseudo point clouds obtained from the stereo camera. In order to obtain high-quality depth maps and point clouds, we first perform various preprocessing and postprocessing on them, such as Speckle filter, probabilistic denoising, etc. Then we propose a multi-scale method for the problem of different ground heights and different scales of point cloud clustering. In addition, we also utilize an RGB monocular camera aligned with a stereo camera for late sensor fusion, which enables partially unlabeled point cloud obstacles to be assigned to a category. Our experimental results show that the proposed system achieves superior results on mobility scooters and electric wheelchairs without requiring a lot of computing resources and equipment costs.