Accuracy assessment of point cloud models generated from mobile phone photos for as-built surveys in construction

Abstract

This study evaluates the accuracy of a 3D point cloud generated from smartphone images for building condition surveys. The experiment was conducted on Building A2 at the main campus of Ho Chi Minh City University of Technology – VNU-HCM. A total of 77 high-resolution images captured with a Xiaomi 12 smartphone were processed using Agisoft Metashape software to reconstruct a dense point cloud, yielding over 81 million points. Ten ground control points (GCPs) were marked on the building surface and their coordinates measured using a total station. Five of these GCPs were used for model georeferencing. Accuracy was assessed based on: (1) the positional errors of the GCPs and (2) the dimensional errors of 17 building edges. After aligning the model using five GCPs, the point cloud was georeferenced and scaled to its real-world dimensions. The mean positional errors of the GCPs were 0.0126 m (X), 0.0108 m (Y), and 0.0151 m (Z), with an overall RMSE of approximately 2.3 cm. The absolute errors in edge length ranged from 0.011 m to 0.055 m, corresponding to relative errors between 0.9% and 4.6%. The findings indicate that smartphone photogrammetry combined with commercial software can achieve sufficient accuracy for as-built surveys, particularly in scenarios with limited access to specialized equipment, tight budgets, and the need for flexible deployment.