Scholarlink

<p>Recent advances in deep learning have enabled accurate video anomaly detection for traffic monitoring and aerial surveillance. However, robust detection in UAV videos captured under adverse weather remains challenging due to severe visual degradations and domain shifts. This paper presents a systematic robustness study of six representative anomaly detection methods—Future Frame Prediction, Spatio-Temporal Dissociation, MNAD, MLEP, ANDT, and ASTT—on two UAV traffic datasets (UIT-ADrone and Drone-Anomaly). To evaluate cross-weather generalization in a controlled manner, we construct adverse-weather variants using established image-to-image translation models for fog, rain, and snow. We report frame-level performance using ROC-AUC and Equal Error Rate (EER) under (i) cross-weather testing and (ii) cross-dataset transfer settings. Results consistently show notable degradation across all models under adverse weather; CNN-based approaches tend to be more resilient than Transformer-based ones under heavy visibility loss and noise patterns. Our findings highlight failure modes that commonly arise in adverse weather and provide practical insights for designing more robust UAV-based traffic anomaly detection systems.</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

Robustness of Traffic Anomaly Detection in UAV Videos: Cross–Weather Generalization

  • 期刊论文

  • 发表日期：

    2026年4月30日

查看全文
  • Hiep T. Nguyen,
  • Khang Nguyen

摘要

Recent advances in deep learning have enabled accurate video anomaly detection for traffic monitoring and aerial surveillance. However, robust detection in UAV videos captured under adverse weather remains challenging due to severe visual degradations and domain shifts. This paper presents a systematic robustness study of six representative anomaly detection methods—Future Frame Prediction, Spatio-Temporal Dissociation, MNAD, MLEP, ANDT, and ASTT—on two UAV traffic datasets (UIT-ADrone and Drone-Anomaly). To evaluate cross-weather generalization in a controlled manner, we construct adverse-weather variants using established image-to-image translation models for fog, rain, and snow. We report frame-level performance using ROC-AUC and Equal Error Rate (EER) under (i) cross-weather testing and (ii) cross-dataset transfer settings. Results consistently show notable degradation across all models under adverse weather; CNN-based approaches tend to be more resilient than Transformer-based ones under heavy visibility loss and noise patterns. Our findings highlight failure modes that commonly arise in adverse weather and provide practical insights for designing more robust UAV-based traffic anomaly detection systems.