Indirect Vision-Based Localization of Cutter Bolts for Shield Machine Cutter Changing Robots.

In operations involving the replacement of shield machine disc cutters, challenges such as limited space, poor lighting, and slurry contamination frequently lead to occlusions and incomplete data when using direct point cloud-based localization for disc cutter bolts. To overcome these issues, this study introduces an indirect visual localization technique for bolts that utilizes image-point cloud fusion. Initially, an SCMamba-YOLO instance segmentation model is developed to extract feature surface masks from the cutterbox. This model, trained on the self-constructed HCB-Dataset, delivers a mAP50 of 90.7% and a mAP50-95 of 82.2%, which indicates a strong balance between its accuracy and real-time performance. Following this, a non-overlapping point cloud registration framework that integrates image and point cloud data is established. By linking dual-camera coordinate systems and applying filtering through feature surface masks, essential corner coordinates are identified for pose calibration, allowing for the estimation of the three-dimensional coordinates of the bolts. Experimental results demonstrate that the proposed method achieves a localization error of less than 2 mm in both ideal and simulated tunnel environments, significantly enhancing stability in low-overlap and complex settings. This approach offers a viable technical foundation for the precise operation of shield disc cutter changing robots and the intelligent advancement of tunnel boring equipment.