Scholarlink

<p>Automated anomaly detection in surveillance videos is crucial for ensuring public safety by reducing human intervention and improving decision-making accuracy. This study proposes SurveillanceNet, a novel two-stream spatio-temporal anomaly recognition framework that efficiently extracts spatial and temporal features for multi-class anomaly classification. The spatial stream utilizes DenseNet201 to extract features from individual RGB frames, while the temporal stream employs SG3I images to capture motion patterns without the computational overhead of optical flow or 3D CNNs. The extracted features are fused and processed by a stacked LSTM for sequence classification. Extensive experiments on the UCF Crime dataset demonstrate the effectiveness of the proposed approach, achieving an AUC of 88.0%, outperforming several CNN-LSTM and C3D-based methods while remaining competitive with transformer-based architectures. Unlike recent deep learning models that rely on memory-intensive transformers or computationally expensive optical flow, SurveillanceNet strikes a balance between accuracy and efficiency, making it suitable for real-time surveillance applications. The results highlight the model’s robustness in capturing spatio-temporal dependencies while addressing inter-class similarity challenges. Future research will focus on enhancing dataset diversity, incorporating advanced attention mechanisms, and exploring hybrid architectures for improved anomaly detection performance.</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

SurveillanceNet: Spatio-temporal anomaly identification in surveillance videos using two-stream CNN and LSTM

  • 期刊论文

  • 发表日期：

    2025年6月16日

查看全文
  • Muhammad Salman Ghauri,
  • Usama Ijaz Bajwa,
  • Gulshan Saleem,
  • Rana Hammad Raza,
  • Muhammad Waqas Anwar

摘要

Automated anomaly detection in surveillance videos is crucial for ensuring public safety by reducing human intervention and improving decision-making accuracy. This study proposes SurveillanceNet, a novel two-stream spatio-temporal anomaly recognition framework that efficiently extracts spatial and temporal features for multi-class anomaly classification. The spatial stream utilizes DenseNet201 to extract features from individual RGB frames, while the temporal stream employs SG3I images to capture motion patterns without the computational overhead of optical flow or 3D CNNs. The extracted features are fused and processed by a stacked LSTM for sequence classification. Extensive experiments on the UCF Crime dataset demonstrate the effectiveness of the proposed approach, achieving an AUC of 88.0%, outperforming several CNN-LSTM and C3D-based methods while remaining competitive with transformer-based architectures. Unlike recent deep learning models that rely on memory-intensive transformers or computationally expensive optical flow, SurveillanceNet strikes a balance between accuracy and efficiency, making it suitable for real-time surveillance applications. The results highlight the model’s robustness in capturing spatio-temporal dependencies while addressing inter-class similarity challenges. Future research will focus on enhancing dataset diversity, incorporating advanced attention mechanisms, and exploring hybrid architectures for improved anomaly detection performance.