Scholarlink

<p>Multi-object tracking has achieved significant success in clear weather conditions due to advancements in deep learning. However, tracking performance decreases significantly under adverse weather conditions such as heavy snow. The lack of large-scale annotated video datasets for these dynamic environments is a major bottleneck in the field. To address this challenge from a computer graphics perspective, we propose SnowTrack-Synth, a physics-based video rendering pipeline that synthesizes realistic dynamic snow and atmospheric fog effects onto clear weather video sequences. By modeling snowflakes as a continuous dynamic particle system that combines kinematics, motion blur, and depth of field effects, our pipeline maintains strict temporal consistency and avoids flickering artifacts common in traditional frame-by-frame noise enhancement. Furthermore, we use these synthetic datasets to train a novel end-to-end framework called WeatherTrack, which integrates a weather-restoration module with a tracking by detection paradigm. By jointly optimizing restoration and detection losses, the network extracts features beneficial for tracking instead of focusing only on human visual quality metrics like PSNR or SSIM. Extensive experiments on synthetic versions of MOT17 and KITTI datasets, along with validation on real-world snowy traffic surveillance videos, show that models trained on SnowTrack-Synth achieve significant improvements in Higher Order Tracking Accuracy by 5.6–14.7% under adverse weather and demonstrate strong cross-domain generalization.</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

Physics-driven snowy scene rendering and robust vehicle tracking under adverse weather

  • 期刊论文

  • 发表日期：

    2026年7月31日

查看全文
  • QiMu Su,
  • Gang Zhou,
  • Yajun Liu,
  • YanYun Zhou,
  • Li Zhang

摘要

Multi-object tracking has achieved significant success in clear weather conditions due to advancements in deep learning. However, tracking performance decreases significantly under adverse weather conditions such as heavy snow. The lack of large-scale annotated video datasets for these dynamic environments is a major bottleneck in the field. To address this challenge from a computer graphics perspective, we propose SnowTrack-Synth, a physics-based video rendering pipeline that synthesizes realistic dynamic snow and atmospheric fog effects onto clear weather video sequences. By modeling snowflakes as a continuous dynamic particle system that combines kinematics, motion blur, and depth of field effects, our pipeline maintains strict temporal consistency and avoids flickering artifacts common in traditional frame-by-frame noise enhancement. Furthermore, we use these synthetic datasets to train a novel end-to-end framework called WeatherTrack, which integrates a weather-restoration module with a tracking by detection paradigm. By jointly optimizing restoration and detection losses, the network extracts features beneficial for tracking instead of focusing only on human visual quality metrics like PSNR or SSIM. Extensive experiments on synthetic versions of MOT17 and KITTI datasets, along with validation on real-world snowy traffic surveillance videos, show that models trained on SnowTrack-Synth achieve significant improvements in Higher Order Tracking Accuracy by 5.6–14.7% under adverse weather and demonstrate strong cross-domain generalization.