Point cloud registration in BIM models for robot localization

Abstract

As digital tools find increasing use in the construction industry, Building Information Modeling (BIM) is used for an increasing amount of tasks. BIM models contain acombination of 3D and semantic information that can be applied for construction planning, progress evaluation, quality control and documentation purposes. In manycases, the BIM model needs to be compared to LiDAR scans of the building, but this process requires manual registration of the gathered data to the correct location in theBIM model.This thesis introduces a novel registration algorithm for localizing point clouds in a BIM model automatically. The registration algorithm uses voxelization, normal alignment, anovel normal filtered template matching approach and the iterative closest point (ICP) algorithm to reliably register the point cloud to the BIM model.For applying the algorithm, a modular software system was developed to gather pointclouds using a mobile LiDAR and Simultaneous Localization and Mapping (SLAM).A 3D visualization shows the live position of the LiDAR, the current point cloud of the enironment and the BIM model at the same time. The viewer includes features for evaluating matching performance and comparing the BIM model to the real environment.We also developed a hand held setup and a robot carried setup to test the system and its possible deployment.The performance of the registration algorithm was evaluated in different environments.We gathered ground truth data to determine the registration accuracy of the algorithm.Our method achieves state-of-the-art accuracy and reliability. In realistic environments,our registration algorithm reached an average accuracy of 15.24 cm and a reliability of more than 96%, while being accurate to up to 3.7 cm in a simpler individual room.We also discuss the influence of different environments and setups on the registration performance.The algorithm overcomes different limitations of other approaches and is suitable tobe used by autonomous systems. Our registration method could also be applied to different use cases in construction and other fields such as archaeology or geology, where environments are scanned and compared to a 3D model.