Quadtree sampling-based superpixels for 3D range data

3D range sensors are currently being used in various fields. Creating 3D range sensors requires various techniques, such as object detection, tracking, classification, 3D SLAM, etc. For the pre-processing step, superpixels can improve the performance of these techniques. This paper proposes a novel over-segmentation algorithm, known as superpixels, for 3D outdoor urban range data. Superpixels are generated with three steps: boundary extraction using a surface change score and sensor models, initial cluster seeding using a quadtree decomposition, and iterative clustering, which adapts a k-means clustering approach with limited search size in the quadtree dimension. The proposed algorithm produces adaptive superpixel sizes that take into account surface and object border information. This reduces memory size more than regular grid methods and represents small objects well with adaptable pixel sizes. The algorithm is verified using the publicly available Velodyne dataset and the manually annotated ground truth. A comparison with the conventional algorithm is also presented.