Analysis of Artificial Intelligence-Based Photogrammetry for Calculating the Volume of Bulk Material Stockpiles

Abstract

This paper presents an automated UAV-based photogrammetric workflow for efficiently and accurately estimating bulk material stockpile volumes, addressing the limitations of conventional manual and LiDAR-based methods. The proposed approach converts UAV video data captured with a 40 MP RGB drone into georeferenced still frames, followed by SIFT and ORB feature extraction and exhaustive matching within COLMAP database. Incremental Structure-from-Motion with bundle adjustment reconstructs a sparse point cloud of 119,424 points and optimized camera parameters, while PatchMatch-based Multi-View Stereo generates a dense cloud of 2.3 million points at a ground sampling distance (GSD) of 0.1 cm. Ground Control Points obtained with RTK-GNSS ensure sub-2 cm georeferencing accuracy. Stockpile volumes are estimated using angle-of-repose height calculations, truncated-pyramid contour integration, and voxel occupancy methods, achieving volume errors of less than 3%. Validation against GPS and terrestrial laser scanning (TLS) references indicates horizontal accuracy of CE90 = 0.208 m, vertical accuracy of LE90 = 0.056 m, and mean reprojection error of 0.19 pixels. The entire process requires only 24 minutes for 199 images, confirming its applicability for industrial monitoring. Overall, the proposed AI-assisted photogrammetric pipeline provides a robust, reproducible, and cost-effective solution for automated stockpile volume measurement, enhancing safety, accuracy, and material management efficiency.