Scholarlink

<p>Surveillance systems generate large volumes of video data, making fast and reliable anomaly detection essential for public-safety and smart-city applications. Existing video anomaly detection (VAD) methods often struggle with imbalanced anomaly occurrence, domain shifts, noise interference, and deployment inefficiency. This paper proposes a hybrid CNN–SRU-LSTM–MIL framework that integrates spatio–temporal CNN features, an annealed top-<i>k</i> MIL ranking loss, efficient SRU-based temporal modeling, and LSTM reconstruction to enhance accuracy, robustness, and runtime performance. To support real-world deployment, we incorporate pruning, quantization, knowledge distillation, and corruption-aware preprocessing. Experiments on UCF-Crime, CUHK Avenue, ShanghaiTech, UMN, and a real-world traffic dataset show consistent improvements, including <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\approx 2\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>≈</mo> <mn>2</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> AUC gains over state-of-the-art baselines and real-time throughput (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\approx 24\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>≈</mo> <mn>24</mn> </mrow> </math></EquationSource> </InlineEquation> FPS on RTX GPUs; 10–14 FPS on Jetson Xavier NX after pruning and 8-bit quantization). These results demonstrate that the proposed framework is accurate, efficient, and suitable for practical intelligent surveillance applications.</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 CNN–SRU/LSTM with multiple instance learning for real-time video anomaly detection in surveillance

  • 期刊论文

  • 发表日期：

    2026年1月27日

查看全文
  • Rajat Gupta,
  • Nidhi Tyagi

摘要

Surveillance systems generate large volumes of video data, making fast and reliable anomaly detection essential for public-safety and smart-city applications. Existing video anomaly detection (VAD) methods often struggle with imbalanced anomaly occurrence, domain shifts, noise interference, and deployment inefficiency. This paper proposes a hybrid CNN–SRU-LSTM–MIL framework that integrates spatio–temporal CNN features, an annealed top-k MIL ranking loss, efficient SRU-based temporal modeling, and LSTM reconstruction to enhance accuracy, robustness, and runtime performance. To support real-world deployment, we incorporate pruning, quantization, knowledge distillation, and corruption-aware preprocessing. Experiments on UCF-Crime, CUHK Avenue, ShanghaiTech, UMN, and a real-world traffic dataset show consistent improvements, including \(\approx 2\%\) ≈ 2 % AUC gains over state-of-the-art baselines and real-time throughput ( \(\approx 24\) ≈ 24 FPS on RTX GPUs; 10–14 FPS on Jetson Xavier NX after pruning and 8-bit quantization). These results demonstrate that the proposed framework is accurate, efficient, and suitable for practical intelligent surveillance applications.