Binary-descriptor-based point cloud registration for open-pit benches

Abstract In digital acceptance of open-pit benches, the accuracy and efficiency of point cloud registration directly affect both workflow throughput and the reliability of excavation-stripping volume assessment. Conventional registration pipelines often suffer from degraded performance in mining scenarios due to ambiguous local geometry, limited overlap between scans, and the large scale of point cloud data. To address these challenges, this paper proposes MineBin-iterative closest point (ICP), a binary-descriptor-based registration method tailored for open-pit bench point clouds. First, leveraging the geometric stability of bench lines, we construct a multi-weight constrained local reference frame that jointly incorporates distance, curvature, and normal-consistency cues to enhance axis stability. Building upon this reference frame, we design an adaptive binary descriptor (MBD) with multi-view rotation augmentation about the Z -axis, which jointly encodes grid-based depth and polar-contour features on three orthogonal planes ( XY, YZ , and XZ ). Finally, robust registration is achieved through a hierarchical optimisation scheme that integrates adaptive Hamming-distance matching, LMeds-based coarse alignment, and point-to-plane ICP refinement. Experiments on UAV photogrammetric point clouds collected from the Qidashan Iron Mine in Liaoning, China, under scenarios involving large rotations, density discrepancies, and low-overlap conditions, demonstrate that MineBin-ICP achieves an root mean square error of 0.1237–0.1924 m and improves registration efficiency by more than 33.3% compared with representative baselines such as FPFH + ICP and SHOT + ICP. The proposed method also exhibits strong scalability and stability on large-scale point clouds, meeting the requirements of digital acceptance in open-pit benches for high-accuracy and high-efficiency registration.