Beyond Accuracy: A Computational–Robustness Comparison of Point Cloud Registration Algorithms

The digital documentation of cultural heritage increasingly relies on the fusion of multi-source data to overcome the limitations of individual sensors. This study evaluates the performance of four registration algorithms, Iterative Closest Point (ICP), RANSAC, Fast Global Registration (FGR), and FilterReg, applied to the fusion of Terrestrial Laser Scanning (TLS) and UAV-based Structure-from- Motion (SfM) point clouds of the Engenho Central do Bracuhy ruins in Brazil. Given the challenging nature of the dataset, characterized by non-uniform density and inherent SfM noise, we utilized the Multimetric Computational-Prediction Efficiency Index (MCPEI) to jointly assess accuracy and processing time. Results indicate that probabilistic methods, specifically FilterReg, outperform geometric-based approaches in multimodal scenarios. FilterReg demonstrated superior performance to Gaussian noise and outliers, maintaining high efficiency scores (> 0.85) where FGR failed to converge. Furthermore, it exhibited rotation invariance without requiring coarse initialization. These findings suggest that probabilistic filtering is a more suitable paradigm for the automated documentation of complex heritage sites than traditional deterministic optimization.