Superpixel Clustering and Planar Fit Segmentation of 3D LIDAR Point Clouds

Terrestrial laser scanning (TLS, also called ground based Light Detection and Ranging, LIDAR) is an effective data acquisition method capable of high precision, detailed 3D models for surveying natural environments. However, despite the high density, and quality, of the data itself, the data acquired contains no direct intelligence necessary for further modeling and analysis - merely the 3D geometry (XYZ), 3-component color (RGB), and laser return signal strength (I) for each point. One common task for LIDAR data processing is the selection of an appropriate methodology for the extraction of geometric features from the irregularly distributed point clouds. Such recognition schemes must accomplish both segmentation and classification. Planar (or other geometrically primitive) feature extraction is a common method for point cloud segmentation, however, current algorithms are computationally expensive and often do not utilize color or intensity information. In this paper we present an efficient algorithm, that takes advantage of both colorimetric and geometric data as input and consists of three principal steps to accomplish a more flexible form of feature extraction. First, we employ a Simple Linear Iterative Clustering (SLIC) super pixel algorithm for clustering and dividing the colorimetric data. Second, we use a plane-fitting technique on each significantly smaller cluster to produce a set of normal vectors corresponding to each super pixel. Last, we utilize a Least Squares Multi-class Support Vector Machine (LSMSVM) to classify each cluster as either "ground", "wall", or "natural feature". Despite the challenging problems presented by the occlusion of features during data acquisition, our method effectively generates accurate (>85%) segmentation results by utilizing the color space information, in addition to the standard geometry, during segmentation.