A skip feature enhanced multi-source fusion framework for switch state detection

Sumin Zhang, Ri Bai, Rui He, Zhiwei Meng and Yongshuai Zhi

International Journal of Rail Transportation, 2025, vol. 13, issue 4, 795-815

Abstract: Railway switches are vital for ensuring the safety of railroad transportation by managing the transitions between different tracks. We propose a skip feature enhanced multi-source fusion network with an attention mechanism for effectively identifying switch states. To further improve the network performance, we propose a data augmentation method of region horizontal splice with 16 images and a novel intersection over union loss function that takes into account the area and aspect ratio of both the predicted and ground truth boxes. Finally, a switch state dataset is built to test our model performance. With a default input size of 640$ \times $×640 pixels, the proposed method achieves a precision of 87.1%, recall of 86.7%, and mAP of 89.4% which is 1.9%, 4.3%, and 1.8% better than Yolov5 baseline respectively. In addition, the inference speed of our proposed method in RTX 3070 is 50 FPS, which meets the requirements of real-time detection.

Date: 2025
References: Add references at CitEc
Citations:

Downloads: (external link)
http://hdl.handle.net/10.1080/23248378.2024.2372729 (text/html)
Access to full text is restricted to subscribers.

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:taf:tjrtxx:v:13:y:2025:i:4:p:795-815

Ordering information: This journal article can be ordered from
http://www.tandfonline.com/pricing/journal/tjrt20

DOI: 10.1080/23248378.2024.2372729

Access Statistics for this article

International Journal of Rail Transportation is currently edited by Wanming Zhai and Kelvin C. P. Wang

More articles in International Journal of Rail Transportation from Taylor & Francis Journals
Bibliographic data for series maintained by Chris Longhurst ().