Approximate triangulation and region growing for efficient segmentation and smoothing of range images

• We directly deduce an approximate surface reconstruction from organized depth data.
• The resulting mesh is used for fast normal computation and caching neighborhoods.
• For smoothing points and normals we present an edge-preserving multilateral filter.
• The smoothed mesh is segmented into planes and other primitives using region growing.
• Experiments show state-of-the-art performance while being significantly faster.

Decomposing sensory measurements into coherent parts is a fundamental prerequisite for scene understanding that is required for solving complex tasks, e.g., in the field of mobile manipulation. In this article, we describe methods for efficient segmentation of range images and organized point clouds. In order to achieve real-time performance in complex environments, we focus our approach on simple but robust solutions. We present a fast approach to surface reconstruction in range images and organized point clouds by means of approximate polygonal meshing. The obtained local surface information and neighborhoods are then used to (1) smooth the underlying measurements, and (2) segment the image into planar regions and other geometric primitives. A comparative evaluation using publicly available data sets shows that our approach achieves state-of-the-art performance while being significantly faster than other methods.