A 3D Multimodal Feature for Infrastructure Anomaly Detection

Yixiong Jing, Brian Sheil

Abstract: Ageing structures require periodic inspection to identify structural defects. However, existing approaches that rely solely on geometric distortions in point clouds often struggle to detect small cracks and non-geometric anomalies. To overcome these limitations, a 3D multimodal feature, 3DMulti-FPFHI, was developed by fusing Fast Point Feature Histograms (FPFH) with a newly designed 3D intensity feature. Integrated into the PatchCore anomaly detection framework, our method enables accurate defect detection directly on two pairs of point clouds without extensive training. An enhanced synthetic dataset, incorporating realistic crack-induced surfaces and intensity variations, is developed to bridge the gap between simulations and real-world scenarios. Validation on both the synthetic dataset and real-world scans from a masonry arch bridge and a full-scale tunnel demonstrates that 3DMulti-FPFHI improves crack detection and identifies intensity anomalies such as water ingress. The method outperforms both FPFH and a state-of-the-art pseudo-multimodal baseline, showing robustness, generalizability, and minimal data requirements.