Open Access

ARTICLE

Automatic Potential Safety Hazard Detection for High-Speed Railroad Surrounding Environment Using Lightweight Hybrid Dual Tasks Architecture

Zheda Zhao, Tao Xu, Tong Yang, Yunpeng Wu^*, Fengxiang Guo^*

The Faculty of Transportation Engineering, Kunming University of Science and Technology, Kunming, 650031, China

* Corresponding Authors: Yunpeng Wu. Email: ; Fengxiang Guo. Email:

(This article belongs to the Special Issue: AI-Enhanced Low-Altitude Technology Applications in Structural Integrity Evaluation and Safety Management of Transportation Infrastructure Systems)

Structural Durability & Health Monitoring 2025, 19(6), 1457-1472. https://doi.org/10.32604/sdhm.2025.069611

Received 26 June 2025; Accepted 13 August 2025; Issue published 17 November 2025

Abstract

Utilizing unmanned aerial vehicle (UAV) photography to timely detect and evaluate potential safety hazards (PSHs) around high-speed rail has great potential to complement and reform the existing manual inspections by providing better overhead views and mitigating safety issues. However, UAV inspections based on manual interpretation, which heavily rely on the experience, attention, and judgment of human inspectors, still inevitably suffer from subjectivity and inaccuracy. To address this issue, this study proposes a lightweight hybrid learning algorithm named HDTA (hybrid dual tasks architecture) to automatically and efficiently detect the PSHs of UAV imagery. First, this HDTA architecture seamlessly integrates both detection and segmentation branches within a unified framework. This design enables the model to simultaneously perform PSH detection and railroad parsing, thereby providing comprehensive scene understanding. Such joint learning also lays the foundation for PSH assessment tasks. Second, an innovative lightweight backbone based on the shuffle selective state space model (S⁴M) is incorporated into HDTA. The state space model approach allows for global contextual information extraction while maintaining linear computational complexity. Furthermore, the incorporation of shuffle operation facilitates more efficient information flow across feature dimensions, enhancing both feature representation and fusion capabilities. Finally, extensive experiments conducted on a railroad environment dataset constructed from UAV imagery demonstrate that the proposed method achieves high detection accuracy while maintaining efficiency and practicality.

---

Keywords

---