Automated Scan-vs-BIM Registration Using Columns Segmented by Deep Learning for Construction Progress Monitoring

Abstract. In construction automation applications, coarse registration between 3D Building Information Modelling (BIM) and the as-built point cloud is vital for the monitoring of construction progress. This can be achieved by extracting highly distinct geometric features in both datasets to speed up the correspondence search. However, the existing geometric feature-based coarse registration methods have limitations in the Architecture, Engineering, Construction & Facility Management (AEC/FM) context because building designs often contain a considerable self-similarity, symmetry, and lack of texture.

In this work, we propose an automatic coarse registration method that is motivated by the Random Sample Consensus (RANSAC) algorithm to estimate the transformation parameters that best align the as-built point cloud in the coordinate frame of the BIM model by matching the corresponding columns. The method is based on the extraction of columns from the as-built point cloud and the as-planned BIM model. For the point cloud data, fully automated column extraction techniques are used by applying deep learning, whereas the BIM model columns are extracted from the available semantic information. Experiments are carried out on real-life datasets from the building construction site to validate the proposed method. The results show that our proposed column-based registration method achieved an RMSE of 2 centimeters , and the cloud-to-cloud mean distance of 1.6cm ± 1.8cm after fine registration. The accuracy of the co-registration result shows that our proposed approach contributes to automating the registration between the as-built point cloud and the as-planned BIM model for construction progress monitoring.