3D GEOMETRIC MODELLING AND POINT CLOUD PROCESSING FOR BUILDING SCENE SEMANTIC SEGMENTATION

Abstract

This dissertation aims to critically evaluate how the precision and accuracy of the Scan to BIM process can affect point cloud processing, the manual reconstruction of an as-built 3D BIM model and semantic segmentation of the building scene. The Confusion Matrix was employed for the processing of the point cloud measurements as well as the 3D BIM model measurements. The on-site measurements taken of the School of Design and Environment 2 Executive Room 1 classroom in the National University of Singapore is taken as the baseline and is assumed to be the ground truth. The tolerance range for the difference in the classroom dimensions is ±14mm.

The results of the five measures - sensitivity, specificity, positive predictive value, negative predictive value and accuracy, were analysed. The sensitivity measure was calculated to be 0.5, the specificity measure to be 1, the PPV measure to be 1, NPV measure to be 0.722 and the accuracy measure to be 0.783. The point cloud was deemed to be moderately accurate in the reconstruction of the 3D BIM model after comparisons to on-site measurements were made. Recommendations provided based on the evaluation of the reconstructed 3D BIM model could be a point of reference for future studies relating to Scan to BIM processes.