A Pairwise-Cascade Distributed 3D Reconstruction Method for Multi-View Industrial Robot Scanning

Multi-view industrial robot scanning requires efficient reconstruction of large point clouds while controlling onboard memory use and maintaining geometric and stitching quality. This paper presents a pairwise-cascade distributed reconstruction framework that addresses serial processing dependencies and onboard memory pressure through layer-wise parallel pairwise merging on edge nodes. Within each merge, octree pruning reduces correspondence search costs, and Voronoi-region classification determines closest points on candidate triangles for point-to-plane iterative closest point (ICP) registration. Distance-weighted fusion consolidates redundant samples in overlapping regions and controls the fusion transition. Dynamic programming optimizes the boundary correspondence path to reduce reliance on local greedy stitching decisions. Boundary interpolation addresses uneven boundary sampling, while edge-length and normal-deviation checks reject invalid candidate faces and trigger local re-solving. Experiments on three industrial components show mean cloud-to-mesh (C2M) distances of 1.88–2.72 mm relative to a centralized reconstruction baseline and stitching anomaly rates of 4.1% at most. The distributed speedup reaches 4.57× on two edge nodes and 7.80× on four nodes, while onboard memory stays within 0.5–0.8 GB for the tested distributed workloads containing 14–34 frames. On the free-form component, the cumulative C2M increase over six merging layers is 0.74 mm. These results demonstrate the framework’s efficiency and onboard memory benefits while characterizing its geometric agreement with centralized reconstruction.