Multi-Scale-Slope Linear Transformer-Based Skeleton Detection for Autonomous Railway Inspection AAV Navigation

Railway inspection is vital to guarantee the safety of railway transportation but now it heavily relies on manual labor, which is difficult and risky in certain regions. The primary premise of autonomous railway inspection using autonomous aerial vehicles (AAVs), with the aim to mitigate the dependency on labor, is achieving autonomous flight along the railway. To ensure the AAV aligns with the railway in high precision, vision-based navigation is more applicable and effective than GNSS-based navigation. In this article, we propose a novel end-to-end network, termed multi-scale-slope linear transformer (MSSLFormer), to detect the railway skeleton directly as target flight trajectory. Our network roots in multi-scale-slope linear attention for enhanced railway skeleton detection and adopts a concurrent head, in which a semantic segmentation branch is included for subsequent inspection tasks. To ensure real-time detection, we leverage attention parameters sharing mechanism when calculating the query-key matrix. Experiments show that our model achieves superior performance, F-score of 86.44%, for skeleton detection on our railway dataset, which contains 9210 labeled images. Relying on our flight strategy, AAV can fly autonomously along the railway up to 15 m/s, demonstrating the great potential for future railway inspection tasks.