Automatic Classification of Urban Objects From Mobile Laser Scanning Data

Abstract

Abstract. The study deals with the automatic supervised classification of urban objects from point clouds collected by the vehicle-based Mobile Laser Scanning (MLS) system. A benchmark dataset representing the Technical University of Munich (TUM) City Campus was used. The main contribution of this article is evaluating the performance difference between kNN, cylindrical and spherical local neighborhood relations in point-based classification of an MLS system using local geometric and shape-based features. The Random Forest (RF) classifier was performed for 8 manually marked classes in the benchmark set: artificial terrain, natural terrain, high vegetation, low vegetation, building, hardscape, artifact and vehicle. We reveal that the cylindrical neighborhood with 13 attributes provides an improvement of 5.2% compared to the spherical neighborhood, while the kNN gave almost the same result as the cylindrical neighborhood (0.8% improvement) in the shortest time. Finally, a new feature set was created by combining the most important features obtained from different neighborhood types. As a result, we achieved 96.9% overall accuracy by using 19 significant features obtained from all neighborhood types for the TUM-MLS1 point cloud.