A Field-Deployed Dual-View 3D Perception and Dynamic Spatial Alignment Framework for Unstructured Heavy-Duty Industrial Environments

Automated rock-breaking operations in underground mining and surface quarrying are heavily hindered by severe object self-occlusion and intense vibration-induced sensor drift. To address these challenges, this paper presents a robust dual-view 3D perception and dynamic spatial alignment framework based on a structured “Perception-Registration-Modeling” architecture. Specifically, a symmetric dual-view RGB-D configuration is deployed to mitigate view-dependent occlusions, coupled with a phased point cloud registration strategy. To handle the severe extrinsic parameter drift caused by repetitive mechanical impacts, we propose an infrastructure-anchored dynamic compensation mechanism. By exploiting the invariant geometric features of the structural grizzly grid as spatial priors, the system models drift vectors and adaptively triggers a phased Iterative Closest Point (ICP) refinement with a decaying search radius, achieving fast on-the-fly spatial consistency recovery within a strict 450 ms industrial control loop. Furthermore, a high-fidelity Screened Poisson Reconstruction (SPR) pipeline is developed to transform noisy point clouds into watertight manifold meshes, enabling precision extraction of geometric centroids and surface normals. Field deployments at an operational mining site demonstrate a mesh-based geometric centroid MAE of 3.2 ± 0.8 mm and a surface normal error of 1.45° ± 0.32°. The proposed framework maintains sub-5 mm spatial fidelity under the tested industrial disturbances. Task-level experiments further demonstrate improved impact-targeting reliability and reduced average strike time compared with the single-view configuration. These results support the use of the proposed framework for robotic impact localization and execution; fragmentation efficiency, tool fatigue, and equipment wear were not directly evaluated.