Scholarlink

<p>Video anomaly detection (VAD) is a critical application of online evolutive learning for real-time image processing. Although Skeleton-based normalizing flow models have recently achieved superior performance by modeling the distribution of normal human poses and flagging deviations as anomalies, their practical deployment is hindered by the heavy computational burden of large pose estimators and the non-causal temporal convolutions that require future frames for inference. In this regard, we propose Lightweight Online Real-time Anomaly Detection (LORAD) for continuous VAD on resource-constrained devices, which addresses the heavy computational burden of pose estimators, and the non-causal temporal processing that precludes genuine streaming inference. Specifically, the conventional AlphaPose backbone is replaced by a compact MobileNetV3-based encoder with 1.27&#xa0;M parameters (65<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> reduction), which is trained via knowledge distillation. Moreover, causal temporal convolutions and an online circular buffer are then introduced into the spatio-temporal graph flow, eliminating all future-frame dependencies. The proposed LORAD achieves frame-level AUC of 84.8% and 71.2% on ShanghaiTech and UBnormal benchmarks, respectively, with end-to-end latency of <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\({\sim }\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>∼</mo> </math></EquationSource> </InlineEquation>40&#xa0;ms (skip=2, 25&#xa0;FPS) on a CPU-only platform, which is fully compatible with the strict latency requirements of real-time and online-adaptive deployment.</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

Lorad: lightweight causal pose flow for online real-time video anomaly detection

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Yanni Zhang,
  • Xinyi Zhu,
  • Peixin Zuo,
  • Ting Huang

摘要

Video anomaly detection (VAD) is a critical application of online evolutive learning for real-time image processing. Although Skeleton-based normalizing flow models have recently achieved superior performance by modeling the distribution of normal human poses and flagging deviations as anomalies, their practical deployment is hindered by the heavy computational burden of large pose estimators and the non-causal temporal convolutions that require future frames for inference. In this regard, we propose Lightweight Online Real-time Anomaly Detection (LORAD) for continuous VAD on resource-constrained devices, which addresses the heavy computational burden of pose estimators, and the non-causal temporal processing that precludes genuine streaming inference. Specifically, the conventional AlphaPose backbone is replaced by a compact MobileNetV3-based encoder with 1.27 M parameters (65 \(\times \) × reduction), which is trained via knowledge distillation. Moreover, causal temporal convolutions and an online circular buffer are then introduced into the spatio-temporal graph flow, eliminating all future-frame dependencies. The proposed LORAD achieves frame-level AUC of 84.8% and 71.2% on ShanghaiTech and UBnormal benchmarks, respectively, with end-to-end latency of \({\sim }\) ∼ 40 ms (skip=2, 25 FPS) on a CPU-only platform, which is fully compatible with the strict latency requirements of real-time and online-adaptive deployment.