TripletA-Net: A Deep Learning Model for Automatic Railway Track Extraction from Airborne LiDAR Point Clouds

Abstract

With the rapid expansion of global railway networks, the demand for efficient railway operation and maintenance has grown significantly. This shift has underscored the need for automated and intelligent detection technologies to replace the traditional, labor-intensive methods in railway maintenance. To address the high cost, limited generalizability, and dependency on manual intervention that challenge conventional railway track extraction methods, this article proposes a novel railway track extraction model that is specifically designed for dealing with the airborne light detection and ranging (LiDAR) data, known as TripletA-Net. TripletA-Net enables automatic and precise semantic segmentation of railway track point clouds. It incorporates a triplet attention mechanism to establish dependencies across different point cloud dimensions, adaptively assigning weights to capture both global and local features comprehensively. A weight-scaling strategy is introduced to further enhance the model's focus on track extraction. In order to reduce overfitting, the AdamW optimizer with decoupled weight decay is employed, addressing common issues encountered with small training datasets. Moreover, the intensity characteristics of the LiDAR point cloud are exploited in place of traditional color features, minimizing errors from multisource data matching. Ablation experiments validate the importance of the weight-scaling module and the AdamW optimizer in improving the model's accuracy. The triplet attention mechanism and intensity information contribute to enhanced precision and generalization. Together these optimizations make TripletA-Net highly effective in track extraction, achieving a mean Intersection over Union of 94.36% on our airborne LiDAR track dataset (acquired from two geographically diverse regions, with a total track length of 2700 m), which is more surpassing than benchmark methods such as PointNet++ (87.87% ), RandLA-Net (91.64% ), and Stratified Transformer (89.49% ).