High-Precision Point Cloud Registration for Long-Span Bridges Based on Iterative Closest-Surface Method

Noncontact, high-fidelity data acquisition has enabled terrestrial laser scanning (TLS) to be widely adopted for bridge geometry measurement and condition monitoring. In TLS applications, point cloud registration directly affects data quality and the correctness of subsequent results. For long-span bridges in large-scale scenes, complex geometry and sparse sampling pose challenges to surface-based, data-driven registration methods, and may degrade registration accuracy. A data-driven approach for high-precision point cloud registration, referred to as the Iterative Closest-Surface (IC-Surface) method, is presented in this study. The method extracts neighboring surface patches via a bounding box and applies random sampling-based plane fitting to derive surface features for registration, effectively mitigating the impact of sparse points and outliers in long-span bridges. Regular points are generated on the source patch and projected onto the corresponding target patch to establish high precision correspondences, yielding a stable and accurate transformation. This method effectively overcomes the limitations of the Iterative Closest Point (ICP), which struggles with unreliable correspondences and outliers. Comparative experiments were conducted using synthetic data, large bridge segments, and full-bridge datasets against commonly used registration methods. The results show that the IC-Surface method maintains high accuracy and stability across varying levels of outliers and overlap ratios. In complex scenes, IC Surface achieves higher registration accuracy than both ICP and the sphere target method, with distance errors reduced from 3 mm to 1 mm and inter-plane angle errors reduced from 0.016 rad to 0.009 rad. These findings demonstrate the method’s broad applicability in digital construction and operation and maintenance assessments of long-span bridges.