Scholarlink

<p>The safe operation of railway systems critically depends on accurate and real-time detection of foreign objects intruding into track areas. To address the limitations of traditional manual monitoring methods, which often suffer from delayed response and misjudgment in complex environments, this study proposes CDH-DETR, a lightweight Transformer-based real-time detection model for railway foreign objects. Building upon the RT-DETR model, our model introduces three key structural innovations to achieve an optimal balance between accuracy, speed, and model complexity. The architecture incorporates a ContextGuided Block module to reconstruct the backbone network, substantially reducing parameter count and computational complexity while maintaining feature representation capability. Furthermore, the model employs the Dynamic upsampling (DySample) to replace traditional operations, enhancing detail perception and restoring critical features in complex scenes without increasing computational burden. Additionally, a dedicated High-level Screening-feature Fusion Pyramid Networks (HSFPN) strengthens multi-scale feature fusion and improves detection robustness for targets with significant scale variations. Experimental results on a constructed railway intrusion dataset demonstrate that CDH-DETR achieves 76.7% <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\textrm{mAP}\)</EquationSource> <EquationSource Format="MATHML"><math> <mtext>mAP</mtext> </math></EquationSource> </InlineEquation> and 96.8 FPS on the HT-SD3403 embedded platform with only 14.86 MB parameters. Compared to the original RT-DETR, our model improves <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\textrm{mAP}\)</EquationSource> <EquationSource Format="MATHML"><math> <mtext>mAP</mtext> </math></EquationSource> </InlineEquation> by 2.0% while reducing parameters by 27% and increasing FPS by 21.1. Ablation studies and comparative analyses confirm that each module effectively enhances model performance, and the overall architecture outperforms mainstream YOLO detection algorithms. The model maintains stable detection performance in practical video validation, demonstrating good practicality and robustness suitable for real-world railway safety monitoring applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A lightweight transformer-based framework for real-time foreign object detection in complex railway environments

  • 期刊论文

  • 发表日期：

    2025年11月7日

查看全文
  • Zhe Dong,
  • Qing Yang,
  • HaoLin Chen,
  • Heng Zhou,
  • Dexin Gao

摘要

The safe operation of railway systems critically depends on accurate and real-time detection of foreign objects intruding into track areas. To address the limitations of traditional manual monitoring methods, which often suffer from delayed response and misjudgment in complex environments, this study proposes CDH-DETR, a lightweight Transformer-based real-time detection model for railway foreign objects. Building upon the RT-DETR model, our model introduces three key structural innovations to achieve an optimal balance between accuracy, speed, and model complexity. The architecture incorporates a ContextGuided Block module to reconstruct the backbone network, substantially reducing parameter count and computational complexity while maintaining feature representation capability. Furthermore, the model employs the Dynamic upsampling (DySample) to replace traditional operations, enhancing detail perception and restoring critical features in complex scenes without increasing computational burden. Additionally, a dedicated High-level Screening-feature Fusion Pyramid Networks (HSFPN) strengthens multi-scale feature fusion and improves detection robustness for targets with significant scale variations. Experimental results on a constructed railway intrusion dataset demonstrate that CDH-DETR achieves 76.7% \(\textrm{mAP}\) mAP and 96.8 FPS on the HT-SD3403 embedded platform with only 14.86 MB parameters. Compared to the original RT-DETR, our model improves \(\textrm{mAP}\) mAP by 2.0% while reducing parameters by 27% and increasing FPS by 21.1. Ablation studies and comparative analyses confirm that each module effectively enhances model performance, and the overall architecture outperforms mainstream YOLO detection algorithms. The model maintains stable detection performance in practical video validation, demonstrating good practicality and robustness suitable for real-world railway safety monitoring applications.