Photogrammetry-Based Segmentation and Volume Calculation of Homogeneous Sand Piles

The automation and enhancement of bulk material volume estimation are of critical importance for the efficient management of inventory in resource industries. This paper introduces a novel approach for the automatic segmentation and volume estimation of sand piles at a quartz quarry, which leverages recent advancements in unmanned aerial vehicle (UAV) technology and deep learning. The Mask R-CNN architecture was employed for instance segmentation, and three distinct input configurations were evaluated: an RGB model utilizing color information, an RGBH model incorporating height data, and an RGB-Blend model that merges color and elevation information via alpha blending. In order to evaluate the models, a custom dataset was employed and a series of metrics were utilized, including the Intersection over Union (IoU) and the F1-score. The method effectively addresses the challenges posed by the homogeneous texture of sand piles and overhanging occlusions, providing accurate segmentation and reliable volume estimations. This approach demonstrates the potential for broader application to other homogeneous bulk materials, offering a scalable and automated solution for industrial volume estimation.