Optimization of Scale-Dependent Parameters for TLS-based Crack Detection in Reinforced Concrete Structures validated by Deep Learning

Automated crack detection in reinforced concrete structures using terrestrial laser scanning (TLS) point clouds faces two technical issues: the absence of guidelines for scale-dependent parameter selection and the lack of systematic evaluation of how acquisition geometry constrains downstream learning. This study addresses both gaps through empirical characterization of TLS measurement conditions and deep learning-based validation across three representative architectures, PointNet, DGCNN, and Point Transformer, under a comprehensive ablation study spanning 126 experimental configurations.

Empirical analysis identifies defect-specific optimal neighborhood radii that differ significantly between crack (median: 0.0125 m) and efflorescence (median: 0.0800 m) detection, confirming that a uniform parameter setting is insufficient and that adaptive scale selection grounded in defect geometry is essential. The study further establishes that restricting incident angles below 80° substantially improves measurement fidelity, and that the minimum detectable crack width corresponds to the actual point spacing (2.5 mm) rather than the nominal scanner resolution. These physically grounded findings are incorporated as handcrafted local geometric features fused into each deep learning architecture via early fusion. Point Transformer achieves the highest robustness, maintaining F1 scores above 0.80 across all configurations and reaching a peak of 0.893 (k = 400, full feature set). Handcrafted geometric features provided architecture-dependent benefits, yielding substantial gains for PointNet (+3.8 points) while offering modest improvements for architectures with built-in local feature extraction. RGB information consistently produced the largest performance improvements across all models, whereas intensity contributed negligibly.

These findings establish physically interpretable guidelines that bridge TLS acquisition geometry and deep learning model behavior, providing a scalable framework for reliable point cloud-based infrastructure inspection.