Abstract
To address the inefficiency and unfairness of traditional manual scrap sorting, we propose the application of 3D vision technology for grading in this work. The multi-view 3D reconstruction algorithm achieves an accuracy within 1 mm in both synthetic and real scrap scenes. This level of accuracy meets the requirements for scrap grading. Subsequently, an automated processing workflow in a non-overlapping scrap scenario is investigated, in which a pipeline based on the multi-view reconstruction integrating point cloud segmentation technique is proposed. Four-point cloud clustering segmentation methods, including Euclidean clustering, Kmeans, DBSCAN and Region Grow, are compared, and it is found that the Euclidean-clustering-based point cloud segmentation algorithm provides the best overall trade-off, achieving an mIoU score of 99.35%, while the thickness measurement error is less than 0.5 mm. The workflow suggests improved robustness and reliability compared to using a single 2D image for thickness inference. These results indicate that 3D vision may provide a valuable basis for the future development of scrap grading systems.