盾构渣土资源化输送带工位骨料颗粒在线检测中，为解决大规模点云导致分割效率低、下采样易损失颗粒形貌细节的难题，提出面向在线检测的颗粒点云实例分割方法——颗粒点云高效无损分割算法（Effi-Particle-Seg），以及几何特征近似估计方法。 Effi-Particle-Seg算法将三维点云投影为二值图像，在二维空间完成实例分割，并在投影阶段同步建立像素—点云映射机制，实现二维实例标签对三维点云的无损回映射，从而在无需下采样的前提下获取单颗粒实例点集并保留完整的三维信息。针对相邻或轻触导致投影前景粘连的场景，采用基于距离变换的标记控制分水岭算法降低欠分割风险； 并基于分割结果，结合有向包围盒（OBB）算法与APR算法实现三轴尺寸、体积及针片状、球度等指标的自动化快速估计。结果表明： 1）对包含951 179个点的点云数据，Effi-Particle-Seg算法分割耗时0.77 s，较传统的DBSCAN算法耗时（21.3 s）降低约96%，在保证三维信息完整性的同时，显著提升在线分割效率； 2）在投影前景粘连场景中，基于距离变换的标记控制分水岭算法可有效抑制欠分割，提高相邻颗粒实例分离的可靠性； 3）随机样本的实测体积与点云重建体积相对误差（绝对值）均小于4%，验证了体积等几何参数近似估计的准确性与可行性； 4）在常规计算机条件下，系统可在约1 s内完成40 cm×40 cm视场范围内颗粒实例分割与几何参数输出，满足输送带在线检测的实时性需求。
Large-scale point clouds generated during online inspection of aggregate particles at conveyor belt stations in shield tunneling muck recycling result in low segmentation efficiency, while downsampling often leads to the loss of morphological details. To address these issues, an online-oriented instance segmentation algorithm for particle point clouds (Effi-Particle-Seg) is developed, together with an approximate geometric feature estimation method. The proposed approach projects three-dimensional (3D) point clouds into binary images for two-dimensional (2D) instance segmentation while simultaneously establishing pixel-point correspondences during projection, enabling lossless backprojection of 2D instance labels to the original 3D points. This process generates per-particle instance point sets without downsampling, thereby preserving complete 3D information. For scenarios in which adjacent or lightly touching particles cause foreground adhesion during projection, a distance transform-based marker-controlled watershed algorithm is introduced to reduce the risk of undersegmentation. Based on the segmented instances, automated and efficient estimation of triaxial dimensions, volume, and shape indices (e.g., flakiness and sphericity) is achieved by combining oriented bounding boxes with air-profile recording. The results demonstrate that (1) for a point cloud containing 951 179 points, Effi-Particle-Seg achieves a segmentation time of 0.77 s, reducing runtime by 96% compared with traditional DBSCAN (21.3 s) while preserving full 3D details; (2) in foreground adhesion scenarios, the proposed marker-controlled watershed algorithm effectively suppresses undersegmentation and improves the reliability of separating neighboring particle instances; (3) for randomly selected samples, the absolute relative error between the measured volume and the point cloud-reconstructed volume is below 4%, demonstrating the accuracy and feasibility of the proposed geometric estimation method; and (4) on a standard PC, the system outputs particle-instance segmentation results and geometric parameters for a 40 cm×40 cm field of view in approximately 1 s, meeting the real-time requirements of conveyor belt inspection.
