Scholarlink

<p>Anomaly detection in tropical heritage structures is often constrained by geometry-only point-cloud descriptors that inadequately capture moisture-related radiometric variation, while prior evaluations frequently confound descriptor contributions with detector-specific behaviour and provide limited statistical attribution. This study isolates the contribution of a hybrid geometric–radiometric descriptor through controlled multi-detector validation. We introduce&#xa0;3D-FPFH-Int, extending Fast Point Feature Histogram with local three-dimensional intensity histograms. An explicit ablation study (geometric-only, radiometric-only, hybrid) is conducted across three detectors—PatchCore, Isolation Forest, and kNN—using&#xa0;500 synthetic instances&#xa0;(including adversarial weak-contrast subsets) and three non-overlapping field scan segments&#xa0;(two bridge spans and one tunnel segment) comprising&#xa0;28.9 million points acquired via terrestrial LiDAR. Training and evaluation employ spatially disjoint partitions to prevent data leakage. The statistical protocol includes two-way ANOVA (Descriptor × Detector), Bonferroni-adjusted post-hoc comparisons, bias-corrected and accelerated 95% bootstrap confidence intervals, Cohen’s d with confidence bounds, and post-hoc power analysis (1 − β = 0.82 for moderate interaction effects, η<sup>2</sup> ≥ 0.05). The hybrid achieves a weighted mean F<sub>1</sub> = 0.559 [0.538–0.580], representing a&#xa0;114% relative improvement over FPFH-only (Cohen’s d = 1.38 [1.27–1.49]). The Descriptor × Detector interaction was not statistically significant&#xa0;(p = 0.15, η<sup>2</sup> = 0.02), indicating that the relative ranking of descriptors remains broadly consistent across the evaluated detectors within the tested conditions. Under&#xa0;40% intensity contrast reduction, ΔF<sub>1</sub> remains + 0.294&#xa0;relative to FPFH. Early crack detection (0.3–0.5&#xa0;mm) yields F<sub>1</sub> = 0.158&#xa0;with localization error &lt; 16&#xa0;mm. Moisture-related anomaly detection achieves recall = 0.85 [0.80–0.90]&#xa0;with&#xa0;68% fewer condensation-induced false positives than intensity-only baselines. Performance degradation remains &lt; 12%&#xa0;under ± 20% point-density perturbation and controlled intensity noise. Validation is restricted to masonry and concrete structures in tropical humid environments using terrestrial LiDAR; generalization to other materials, climates, or sensing modalities requires independent verification. Code and synthetic data are publicly available to support reproducibility.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

3D-FPFH-Int: a hybrid geometric–radiometric descriptor for structural surface anomaly detection in tropical heritage

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月23日

查看全文
  • Hassan Gbran,
  • Hendro Murtiono

摘要

Anomaly detection in tropical heritage structures is often constrained by geometry-only point-cloud descriptors that inadequately capture moisture-related radiometric variation, while prior evaluations frequently confound descriptor contributions with detector-specific behaviour and provide limited statistical attribution. This study isolates the contribution of a hybrid geometric–radiometric descriptor through controlled multi-detector validation. We introduce 3D-FPFH-Int, extending Fast Point Feature Histogram with local three-dimensional intensity histograms. An explicit ablation study (geometric-only, radiometric-only, hybrid) is conducted across three detectors—PatchCore, Isolation Forest, and kNN—using 500 synthetic instances (including adversarial weak-contrast subsets) and three non-overlapping field scan segments (two bridge spans and one tunnel segment) comprising 28.9 million points acquired via terrestrial LiDAR. Training and evaluation employ spatially disjoint partitions to prevent data leakage. The statistical protocol includes two-way ANOVA (Descriptor × Detector), Bonferroni-adjusted post-hoc comparisons, bias-corrected and accelerated 95% bootstrap confidence intervals, Cohen’s d with confidence bounds, and post-hoc power analysis (1 − β = 0.82 for moderate interaction effects, η2 ≥ 0.05). The hybrid achieves a weighted mean F1 = 0.559 [0.538–0.580], representing a 114% relative improvement over FPFH-only (Cohen’s d = 1.38 [1.27–1.49]). The Descriptor × Detector interaction was not statistically significant (p = 0.15, η2 = 0.02), indicating that the relative ranking of descriptors remains broadly consistent across the evaluated detectors within the tested conditions. Under 40% intensity contrast reduction, ΔF1 remains + 0.294 relative to FPFH. Early crack detection (0.3–0.5 mm) yields F1 = 0.158 with localization error < 16 mm. Moisture-related anomaly detection achieves recall = 0.85 [0.80–0.90] with 68% fewer condensation-induced false positives than intensity-only baselines. Performance degradation remains < 12% under ± 20% point-density perturbation and controlled intensity noise. Validation is restricted to masonry and concrete structures in tropical humid environments using terrestrial LiDAR; generalization to other materials, climates, or sensing modalities requires independent verification. Code and synthetic data are publicly available to support reproducibility.