Leakage Segmentation Algorithm for Complex Tunnel Scenes Based on LiDAR-Vision Data

To address challenges in railway tunnel leakage detection, such as complex field environments, strong interference in visual images, and limited data representation capability of single-sensor systems, this study proposes a tunnel leakage segmentation algorithm based on LiDAR–Vision collaboration. First, a multi-object tunnel database is constructed by integrating point cloud depth information containing distance features, point cloud intensity information with high reflectivity to leakage water, and high-resolution visual texture information. Second, this study proposes Tunnel-KNet, a semantic segmentation framework for tunnel leakage detection. It incorporates a Swin Transformer encoder to extract robust global features and a multi-decoder architecture to refine feature representations across multiple scales, leading to significantly improved detection performance in complex scenarios. Experimental results show that the model achieves an accuracy of 96.48% and a mean Intersection over Union (mIoU) of 84.19% on the tunnel LiDAR–Vision dataset. This research validates the advantages of LiDAR–Vision data fusion in tunnel leakage segmentation and provides an innovative solution for intelligent monitoring of tunnel defects.