Scholarlink

Road surface condition (RSC) monitoring is essential for enhancing vehicle safety and accident prevention. This study investigates the application of computer vision techniques for real-time sensing of road surface conditions. We introduce a novel dataset named WARD (Weather-Aware Road Dataset), a comprehensive collection of almost \(55\ 000\) images collected in real-world driving scenarios across diverse seasonal and weather conditions, designed to advance RSC detection, now available for download. We thoroughly evaluate state-of-the-art computer vision models, specifically MobileNet and EfficientNet, on both the WARD and publicly available RoadSaW datasets, providing insights into their classification performance. MobileNet exhibited superior classification and inference speed results, processing images at up to 30 fps on an affordable GPU. To improve real-time efficiency, we employ temporal smoothing through moving window aggregation. Our findings validate the potential of non-contact, camera-based RSC monitoring, showcasing its practicality and cost-effectiveness compared to other sensors.

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

WARD: Weather-Aware Road Surface Condition Monitoring Dataset

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Soňa Nesnídalová,
  • Tomáš Kerepecký,
  • Adam Novozámský

摘要

Road surface condition (RSC) monitoring is essential for enhancing vehicle safety and accident prevention. This study investigates the application of computer vision techniques for real-time sensing of road surface conditions. We introduce a novel dataset named WARD (Weather-Aware Road Dataset), a comprehensive collection of almost \(55\ 000\) images collected in real-world driving scenarios across diverse seasonal and weather conditions, designed to advance RSC detection, now available for download. We thoroughly evaluate state-of-the-art computer vision models, specifically MobileNet and EfficientNet, on both the WARD and publicly available RoadSaW datasets, providing insights into their classification performance. MobileNet exhibited superior classification and inference speed results, processing images at up to 30 fps on an affordable GPU. To improve real-time efficiency, we employ temporal smoothing through moving window aggregation. Our findings validate the potential of non-contact, camera-based RSC monitoring, showcasing its practicality and cost-effectiveness compared to other sensors.