Research on bulk grain pile volume measurement method based on DUSt3R stereo deep learning

To address issues such as high equipment costs, reliance on camera calibration, and poor adaptability to complex shapes in traditional bulk grain pile volume measurement methods, this study proposes an intelligent volume estimation method for bulk grain pile volume based on DUSt3R 3D reconstruction point clouds. This method leverages DUSt3R's attention mechanism and dense matching technology to generate 3D point clouds end-to-end without requiring pre-calibrated camera parameters. A point cloud optimization module tailored to the characteristics of grain pile is constructed, which combines statistical filtering with RANSAC plane detection technology to enhance point cloud noise removal capabilities and employs DBSCAN clustering to achieve precise segmentation between the grain pile and the ground surface. Results indicate that this approach not only effectively eliminates the need for camera calibration and significantly improves the accuracy of point cloud noise processing and segmentation, but also adaptively fits complex grain pile morphologies via dynamic grid projection and Alpha Shape surface reconstruction. It ensures measurement accuracy while substantially reducing hardware costs, demonstrating excellent engineering applicability. Validation tests conducted on six typical grain pile morphologies yielded an average measurement error of approximately 5%, with data acquisition achievable using only ordinary cameras. This volume measurement method can be efficiently integrated with grain leveling robot equipment, providing a low-cost, high-precision technical solution for bulk grain pile volume measurement and automated operation guidance.