From clustering to classification: An uncertainty analysis of a classification approach using a Gaussian mixture model

Abstract

This paper presents a systematic pipeline for the segmentation of LiDAR data. The LiDAR data is specifically acquired using the LiDAR sensor of a BLK2GO and stored in a PLY format. The presented pipeline uses common post-processing steps for point cloud processing. First, a RANSAC algorithm is used to remove the indoor point cloud scene from the walls, using plane detection for dominant structures. Subsequently, the indoor scene is clustered by a DBSCAN algorithm. The clustered indoor scene objects are then briefly classified manually. After the manual step, the pipeline is using a Gaussian Mixture Model (GMM). The GMM classifies the objects based on their probabilistic distributions of the geometric dimension. Preliminary results indicate that this workflow enables an efficient object detection with a minimal manual effort. Furthermore, the paper investigates uncertainties of process steps and gives new approaches for dealing with the uncertainties. This pipeline aims to establish a practical approach for generating input data for Scan2BIM applications.