Local Geometric Perception for Climbing Robots on Double-Deck Steel Truss Bridges Using CAD Hole-Layout Priors and Depth-Derived Geometry

Close-range bridge inspection exposes personnel to risks associated with work at height and in confined spaces. Magnetic climbing robots can reduce this exposure, yet dense fastener groups and unsupported regions beyond truss member edges impose local geometric constraints on robot motion. Forward-looking depth observations of these regions are often partial and noisy. This study presents a local geometric perception method that produces two constraints. A computer-aided design (CAD)-constrained branch matches partial rivet observations to hole-layout templates and converts the recovered layout into an inward-offset rivet-passage boundary. A CAD-independent depth branch fits the supporting-surface boundary near a truss member edge. The method was evaluated in simulation, laboratory mock-ups, and a single-case field feasibility study. Under simulation conditions, CAD-constrained recovery reduced lateral mean absolute error by 83.17–85.14% relative to a geometry-only ablation baseline, and all generated passage boundaries retained positive theoretical clearance from the corresponding fastener contours. Laboratory trials achieved template locking; qualitative overlay inspection found no visible boundary intrusion for three rivet layouts. The single-case field study also achieved template locking and generated a passage boundary without visible intrusion into the fastener contours using a manually measured bolt-layout prior. The member-edge branch produced valid fits in 79.59% of field frames, with a mean branch processing time of 23.37 ms. These results support the feasibility of combining CAD hole-layout priors and depth geometry to provide complementary local geometric constraints for subsequent motion planning.