Accurate volume measurement of bulk piles is essential for material management in industrial scenarios. To address the challenges of irregular pile geometries and noisy environments, this paper proposes a high-precision volume estimation method based on 3D point cloud semantic segmentation. The method includes a comprehensive preprocessing pipeline and an improved PointNet++ model with an adaptive pooling mechanism to enhance segmentation robustness under varying point densities. For volume computation, a Delaunay triangulation-based elevation projection model is constructed, effectively reducing computational complexity via dimensionality reduction. Experiments on both standard geometries and simulated bulk piles demonstrate that the proposed method achieves high accuracy (average error < 1%) and efficiency (runtime < 50 s), outperforming traditional convex hull and voxel-based methods. The framework provides a practical and scalable solution for intelligent bulk material volume estimation.