6D DBSCAN-based segmentation of building point clouds for planar object classification

Abstract

Abstract Due to constraints in manufacturing and construction, buildings and many of the manmade objects within them are often rectangular and composed of planar parts. Detection and analysis of planes is, therefore, central to processing point clouds captured in these spaces. This paper presents a study of the semantic information stored in the planar objects of noisy building point clouds. The dataset considered is the Scene Meshes Dataset with aNNotations (SceneNN), a collection of over 100 indoor scenes captured by consumer-grade depth cameras. All planar objects within the dataset are detected using a new point cloud segmentation method that applies Density Based Spatial Clustering of Applications with Noise (DBSCAN) in a six dimensional clustering space. With all planes isolated, an extensive list of features describing the planes is extracted and studied using feature selection. Then dimensionality reduction and unsupervised learning are used to explore the discriminative ability of the final feature set as well as emergent class groupings. Finally, we train a bagged decision tree classifier that achieves 71.2% accuracy in predicting the object class from which individual planes originate.