Extracting Multiple Planar Surfaces Effectively and Efficiently Based on 3D Depth Sensors

Abstract

Reliable and efficient planar surface extraction based on the 3D depth sensors is a crucial component in mobile robotics. However, extracting planes efficiently remains challenging, especially when small planar structures are urged to be perceived. This paper proposes a new method that can extract the multi-scale planar surfaces efficiently. The depth image is dynamically divided into rectangle regions, where points in each region lie on a common plane. Then, the planar primitives are generated by clustering these regions into some distinct groups according to their plane parameters. Finally, the pixel-wise segmentation results are achieved by growing each distinct group. The first novelty is the dynamic region size adjusting algorithm that can reduce the number of regions to be clustered and improve the plane fitting accuracy. The second one is the region clustering algorithm, whose worst-case time complexity is guaranteed to be log-linear. Comprehensive experiments show that our algorithm outperforms the state-of-the-art methods in quality and runs an order of magnitude faster.

Keywords

• Computer vision
• depth sensor
• planar region clustering
• plane extraction
• robot vision systems