Object extraction based on 3d-segmentation of LiDAR data by combining mean shift with normalized cuts: two examples from urban areas

In this work, we have looked into the problem of urban analysis using airborne LiDAR data based on the strategy of classification by segmentation. Segmentation is a key and hard step in the processing of 3D point clouds, which is not perfectly solved in view of different applications. A new 3d segmentation method incorporating the advantages of nonparametric and spectral graph clustering is presented here to facilitate the task of object extraction in urban areas. This integrated method features local detection of arbitrary modes and globally optimized organization of segments concurrently, thereby making it particularly appropriate for partitioning raw airborne LiDAR data of urban areas into segments approximating semantic entities. Two examples in urban areas - flyover and vehicle are chosen as interest objects to be extracted by a classification-based step. The approach has been tested on LiDAR data of dense urban areas, and the results that are obtained have been compared with manual counts and showed us the efficiency and reliability of the strategy.