Scholarlink

<p>With the increasing demand for real-time highway safety monitoring, detecting abnormal vehicle behaviors at long distances remains challenging. Traditional computer vision methods relying on Intersection-over-Union (IoU)-based bounding box variations often fail to distinguish stationary or slow-moving distant vehicles from normal traffic due to minimal short-term geometric changes in surveillance footage. To address this, we propose RoadTrack for real-time long-distance highway parking detection. Our approach employs homography transformation to map 2D image coordinates to real-world road plane coordinates, enabling precise vehicle motion modeling. By combining object detection with multi-object tracking, stable trajectories are constructed and processed through a lane-adaptive Kalman filter to estimate instantaneous velocities. This velocity-based strategy overcomes the limitations of IoU-dependent methods, achieving rapid anomaly identification even for distant targets. Experimental evaluations demonstrate that RoadTrack attains 96% detection accuracy (10% improvement) while reducing detection latency to 3.1 seconds (58% faster). Additionally, to establish performance evaluation benchmark for subsequent research, we introduce the first large-scale dataset for long-distance highway parking detection. It includes 1,000 videos and is available at https://github.com/corfyi/RoadTrack.</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

RoadTrack: Real-Time Distant Parking Detection via Visual Geometric Projection and Speed Estimation

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Kefu Yi,
  • Longbing Luo,
  • Chunlin Xie,
  • Yufan Zhu,
  • Yunyang Tian

摘要

With the increasing demand for real-time highway safety monitoring, detecting abnormal vehicle behaviors at long distances remains challenging. Traditional computer vision methods relying on Intersection-over-Union (IoU)-based bounding box variations often fail to distinguish stationary or slow-moving distant vehicles from normal traffic due to minimal short-term geometric changes in surveillance footage. To address this, we propose RoadTrack for real-time long-distance highway parking detection. Our approach employs homography transformation to map 2D image coordinates to real-world road plane coordinates, enabling precise vehicle motion modeling. By combining object detection with multi-object tracking, stable trajectories are constructed and processed through a lane-adaptive Kalman filter to estimate instantaneous velocities. This velocity-based strategy overcomes the limitations of IoU-dependent methods, achieving rapid anomaly identification even for distant targets. Experimental evaluations demonstrate that RoadTrack attains 96% detection accuracy (10% improvement) while reducing detection latency to 3.1 seconds (58% faster). Additionally, to establish performance evaluation benchmark for subsequent research, we introduce the first large-scale dataset for long-distance highway parking detection. It includes 1,000 videos and is available at https://github.com/corfyi/RoadTrack.