EDT-Net: A Lightweight Tunnel Water Leakage Detection Network Based on LiDAR Point Clouds Intensity Images

Water leakage poses a significant threat to the safe operation of tunnels. We utilized a Mobile Laser Scanner (MLS) to collect point cloud data under adverse tunnel conditions. A data mapping approach was employed to generate MLS point cloud intensity images. Tailored for multiscale point cloud intensity images, we devised a lightweight object detection network to identify areas affected by water leakage promptly. Integrating efficient receptive field expansion convolution (EREC) into lightweight network models facilitated efficient feature extraction. Additionally, we designed an effective attention-inducing downsampling unit (EADU) to construct a tunnel leakage detection model. This module comprehensively handles target features, enhances target context information, enlarges the receptive field, and establishes a unique information processing framework for detecting various multisize targets, achieving outstanding detection performance. Moreover, we developed a dynamic threshold adaptive loss function that automatically adjusts the loss function based on leakage detection performance to enhance the model's ability to detect challenging targets. Finally, we employed a twin attention-guided dynamic detection-head (TADD) to improve detection performance. Experimental results demonstrate that our method effectively transforms the process from MLS point cloud data acquisition to high-precision target detection. The leakage detection network has achieved an optimal balance between efficiency and accuracy, surpassing comparative methods, thereby ensuring the secure operation of shield tunnels.