Scholarlink

The growth of urban cycling brings safety challenges, particularly in detecting unsafe behaviors like excessive swerving, which can lead to collisions or running off the road. Identifying these patterns in video data can improve road safety and aid infrastructure planning. This study introduces a framework for automatically measuring swerving patterns by tracking the distance between landmarks defining the road and the bicycle. By leveraging the Segment Anything Model (SAM) fine-tuned with PerSAM, the method involves segmenting relevant landmarks in video frames to calculate the pixel distance between these landmarks and the bicycle, thus illustrating the swerving patterns over time. The methodology is applied to a dataset comprising 54 videos recorded in diverse cycling environments in Groningen, the Netherlands, categorized into three subdatasets based on the type of bicycle path. This research evaluates three segmentation variants: baseline PerSAM for single-class segmentation, Multi-Class PerSAM (MCP) for 2-class segmentation, and MCP for 3-class segmentation, alongside outlier removal techniques and signal smoothing to enhance measurement accuracy. Quantitative and qualitative results reveal that the optimal configuration involves two landmarks, Isolation Forest for outlier removal, and a rolling average for signal smoothing. This configuration demonstrates a strong correlation with ground truth measurements, offering valuable insights for efficient analysis of cyclists’ swerving behavior in real-world environments.

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

Landmark-Based Cyclist Swerving Analysis via Class-Aware One-Shot Segmentation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Maya Aghaei,
  • Alexander van Meekeren,
  • Frank Westerhuis,
  • Dick de Waard,
  • Klaas Dijkstra

摘要

The growth of urban cycling brings safety challenges, particularly in detecting unsafe behaviors like excessive swerving, which can lead to collisions or running off the road. Identifying these patterns in video data can improve road safety and aid infrastructure planning. This study introduces a framework for automatically measuring swerving patterns by tracking the distance between landmarks defining the road and the bicycle. By leveraging the Segment Anything Model (SAM) fine-tuned with PerSAM, the method involves segmenting relevant landmarks in video frames to calculate the pixel distance between these landmarks and the bicycle, thus illustrating the swerving patterns over time. The methodology is applied to a dataset comprising 54 videos recorded in diverse cycling environments in Groningen, the Netherlands, categorized into three subdatasets based on the type of bicycle path. This research evaluates three segmentation variants: baseline PerSAM for single-class segmentation, Multi-Class PerSAM (MCP) for 2-class segmentation, and MCP for 3-class segmentation, alongside outlier removal techniques and signal smoothing to enhance measurement accuracy. Quantitative and qualitative results reveal that the optimal configuration involves two landmarks, Isolation Forest for outlier removal, and a rolling average for signal smoothing. This configuration demonstrates a strong correlation with ground truth measurements, offering valuable insights for efficient analysis of cyclists’ swerving behavior in real-world environments.