Scholarlink

<p>Video anomaly detection (VAD) is crucial for identifying abnormal behaviors or objects in video sequences. Traditional unsupervised methods, while generalizable, often suffer from high false alarm rates. Supervised methods, though effective for known anomalies, struggle with unknown ones. This study introduces a hybrid supervised framework that leverages a small amount of known abnormal data for supervised learning, ensuring clear boundary distinction between normal and abnormal data, while maintaining generalization ability for unknown anomalies through feature-level unsupervised learning. Experimental results on the UBnormal and ShanghaiTech datasets demonstrate our model’s superior performance compared to existing methods, with significant improvements in generalization ability. We release our code at <a href="https://github.com/waibibabo311/hybrid_supervised_vad_code">https://github.com/waibibabo311/hybrid_supervised_vad_code</a>.</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

Hybrid supervised learning for enhanced open-world video anomaly detection

  • 期刊论文

  • 发表日期：

    2026年4月16日

查看全文
  • Weijie Gao,
  • Jiaxu Leng,
  • Xiangqi Meng,
  • Changhe Tu

摘要

Video anomaly detection (VAD) is crucial for identifying abnormal behaviors or objects in video sequences. Traditional unsupervised methods, while generalizable, often suffer from high false alarm rates. Supervised methods, though effective for known anomalies, struggle with unknown ones. This study introduces a hybrid supervised framework that leverages a small amount of known abnormal data for supervised learning, ensuring clear boundary distinction between normal and abnormal data, while maintaining generalization ability for unknown anomalies through feature-level unsupervised learning. Experimental results on the UBnormal and ShanghaiTech datasets demonstrate our model’s superior performance compared to existing methods, with significant improvements in generalization ability. We release our code at https://github.com/waibibabo311/hybrid_supervised_vad_code.