Scholarlink

<p>Current skeleton-based video anomaly detection (VAD) methods still have limits. They do not fully link local skeleton motion, global temporal context, and node-edge relations. In this paper, “cross-scale” has a clear scope. It covers temporal scales from short clips to long-range dependencies. It also covers spatial scales from local joint neighborhoods to distant node-edge relations. We propose the Cross-Scale Gated Embedding Graph Model for Skeleton-based Anomaly Detection (CSGESAD). Different from STG-NF and HSTGCNN, CSGESAD enhances graph features before probability evaluation. MGIFU fuses local and global features with adaptive gates. AEANet extends graph propagation through node-edge memory and sinusoidal features. Experiments on four datasets show higher AUC with a compact parameter scale.</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

Cross-scale gated embedding graph model for skeleton-based anomaly detection

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Yutong He,
  • Peng Zhang,
  • Tianhuan Huang,
  • Xianye Ben,
  • Lei Chen

摘要

Current skeleton-based video anomaly detection (VAD) methods still have limits. They do not fully link local skeleton motion, global temporal context, and node-edge relations. In this paper, “cross-scale” has a clear scope. It covers temporal scales from short clips to long-range dependencies. It also covers spatial scales from local joint neighborhoods to distant node-edge relations. We propose the Cross-Scale Gated Embedding Graph Model for Skeleton-based Anomaly Detection (CSGESAD). Different from STG-NF and HSTGCNN, CSGESAD enhances graph features before probability evaluation. MGIFU fuses local and global features with adaptive gates. AEANet extends graph propagation through node-edge memory and sinusoidal features. Experiments on four datasets show higher AUC with a compact parameter scale.