| In naturally stacked rock heaps, surface rocks not only occlude one another but also experience inter-layer occlusion with deeper layers, resulting in visual information loss and significant challenges for accurate visual gradation detection. To address these challenges, this study proposes a vision-based intelligent inference framework for rockfill gradation that accounts for occlusion. A visual contour perception model was developed to reconstruct the contours of occluded rocks under natural stacking conditions, extracting their characteristic parameters. Considering the uncertainty in stacking angles, the three-dimensional particle sizes were estimated based on the random sampling of particle shapes and stacking angles, and the visual gradation on the surface layer was evaluated. To address discrepancies between visual and actual gradation caused by inter-layer occlusion, a mapping was constructed using the support vector regression (SVR) algorithm, enabling accurate inference of actual gradation while accounting for occlusion. Field experiments were conducted based on a certain hydropower project, verifying the feasibility and effectiveness of the proposed gradation inference method. The developed software and hardware system was deployed for on-site gradation detection, demonstrating results highly consistent with those obtained from the traditional sieving method. Compared to the conventional method, the proposed system provides significant advantages, including minimal interference, non-contact operation capability, and single-test completion within 1 minute. non-contact operation, and a less than 1 minute detection time per measurement. Furthermore, the testing frequency for transition materials and upstream rockfill was increased by 75 times and 250 times respectively. This system enables large-scale, high-efficiency, and high-precision gradation of rock materials before placement on the dam. |