hierarchical extraction of multiple objects from mobile laser scanning data

This paper proposes an efficient method to extract multiple objects from mobile laser scanning data. The proposed method firstly generates multi-scale supervoxels from 3D point clouds using colors, intensities and spatial distances. Then, a graph-based segmentation method is applied to segment the supervoxels by integrating their colors, intensities, normal vectors, and principal directions. Then, the saliency of each segment is calculated and the most salient segment is selected as a seed to cluster for objects clustering. Hence, the objects are classified and the constraint conditions of object's category are included to re-clustering for more accurate extraction of objects. Experiments show that the proposed method has a promising solution for extracting buildings, ground, street lamps, trees, telegraph poles, traffic signs, cars, enclosures and the objects extraction overall accuracy is 92.3%.