AF-EIDD: an anchor-free, end-to-end obstacle image denoising detection network for rail transit

Abstract With the growth of urban populations and the increasing demand for transportation, urban rail transit plays a crucial role in modern cities. However, adverse weather conditions pose significant challenges to the robustness of obstacle detection for most traditional networks in rail environments. This paper proposes a highly robust anchor-free, end-to-end obstacle image denoising detection network (AF-EIDD) for rail transit. AF-EIDD enhances obstacle detection by denoising images and extracting critical features for accurate obstacle identification, even under adverse weather conditions. The network is lightweight yet capable of capturing local and global image characteristics, ensuring high efficiency and precision. The proposed method has been validated on datasets collected from real-world rail scenarios, including the Nanning Metro Line 2 test route and tunnel environments. AF-EIDD achieves high average mean average precision while maintaining a lightweight design, demonstrating practical applicability for real-time obstacle detection. A dedicated robustness testing dataset was also developed, covering nighttime, heavy rain, and dense fog conditions. Experimental results indicate that AF-EIDD outperforms state-of-the-art obstacle detection networks in terms of robustness, highlighting its potential for deployment in complex rail transit environments.