Scholarlink

<p>Real-time railway perception requires simultaneous understanding of safety-critical targets and structured railway scenarios under strict computational constraints. Existing methods are usually designed for intrusion detection or railway scene understanding independently, or are adapted from general purpose multitask frameworks for road scenarios. In this work, a unified real-time railway multitask learning framework named RailMTL is proposed to jointly perform foreign object intrusion detection, track drivable area segmentation, and track line segmentation. Three key components are introduced to address railway-specific challenges. Multi-Window Global Partition Attention Module is employed to capture fine-grained local details and long-range global context with high efficiency. Cross-Domain Gated Fusion Module is designed to selectively integrate complementary features across heterogeneous task branches while suppressing task interference. The Uncertainty-aware Multitask Training Strategy is further adopted to stabilize joint optimization by explicitly modeling task dependent uncertainty. Experiments on the OSDaR23 and BDD100K datasets demonstrate strong and balanced multitask performance. On OSDaR23, 43.4% mAP is achieved for intrusion detection, together with 88.9% mIoU for track drivable area segmentation and 71.3% mIoU for track line segmentation. On BDD100K, detection and segmentation performance reach 85.3% mAP and 92.8% mIoU, respectively, while real-time inference efficiency is maintained.</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

RailMTL: real-time efficient perception generator for railway multitask learning

  • 期刊论文

  • 发表日期：

    2026年7月8日

查看全文
  • Lu Yang,
  • Xinyi Wang,
  • Guodong Zhu,
  • Mengpan Xiao,
  • Yiwen Bai,
  • Chongke Bi

摘要

Real-time railway perception requires simultaneous understanding of safety-critical targets and structured railway scenarios under strict computational constraints. Existing methods are usually designed for intrusion detection or railway scene understanding independently, or are adapted from general purpose multitask frameworks for road scenarios. In this work, a unified real-time railway multitask learning framework named RailMTL is proposed to jointly perform foreign object intrusion detection, track drivable area segmentation, and track line segmentation. Three key components are introduced to address railway-specific challenges. Multi-Window Global Partition Attention Module is employed to capture fine-grained local details and long-range global context with high efficiency. Cross-Domain Gated Fusion Module is designed to selectively integrate complementary features across heterogeneous task branches while suppressing task interference. The Uncertainty-aware Multitask Training Strategy is further adopted to stabilize joint optimization by explicitly modeling task dependent uncertainty. Experiments on the OSDaR23 and BDD100K datasets demonstrate strong and balanced multitask performance. On OSDaR23, 43.4% mAP is achieved for intrusion detection, together with 88.9% mIoU for track drivable area segmentation and 71.3% mIoU for track line segmentation. On BDD100K, detection and segmentation performance reach 85.3% mAP and 92.8% mIoU, respectively, while real-time inference efficiency is maintained.