Hybrid Deep Learning and Geometric Analysis for Railway Sinkhole Detection in Lidar Point Clouds

This paper presents a hybrid framework for railway sinkhole detection, leveraging deep learning in conjunction with geometric analysis to enhance detection accuracy. A synthetic point cloud dataset was generated, incorporating artificial sinkholes to simulate real-world conditions, following the methodology established in our prior work [1]. The proposed model employs Kernel Point Convolution (KPConv) to directly process raw LiDAR point clouds, effectively capturing both spatial patterns and geometric characteristics of sinkholes. The approach was evaluated on real-world LiDAR-derived Digital Elevation Models (DEMs), demonstrating a substantial improvement in segmentation performance. Post-processing techniques further refined the predictions, increasing precision from 66 % to 73 % and recall from 76 % to 82 %, resulting in a final Intersection over Union (IoU) of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{6 2 \%}$</tex>. These results highlight the benefits of integrating geometric constraints with deep learning to enhance the reliability of sinkhole detection. Furthermore, the proposed method is compatible with Digital Twin frameworks for railway infrastructure, enabling real-time monitoring and automated alert systems for proactive risk mitigation.