Scholarlink

<p>Anomaly detection by video surveillance needs to ensure public safety in metropolitan and sensitive environments. Most solutions today, though, are saddled with accuracy, scalability, and reliability issues due to the use of fixed detection patterns and single-modal information. In a bid to enhance anomaly detection efficiency in real-time, this paper presents Conditional Attention Mechanism Module (CAMM)- Generative Adversarial Network (GAN) (CAMM-GAN), the first real-time context-aware multimodal GAN framework that (i) conditions attention across video, audio and environmental sensors via a novel CAMM, (ii) uses GAN-based generative regularization to model the long-tailed ‘normal’ manifold, and (iii) incorporates reinforcement learning to adapt thresholds and attention online and UCF-Crime while remaining practical for deployment after standard inference optimizations. CAMM-GAN’s hierarchically combines visual, auditory, and environmental sensor data by eliminating redundant features and promoting context awareness. Temporal activity is abstracted in the shape of 3D Convolutional Neural Networks from video, sequential patterns in audio are abstracted by Recurrent Neural Networks, and decision tree models are used for environmental sensor data such as temperature and smoke. These conflicting data sources are combined into a single framework for multi-modal feature integration through the weighted normalization technique. Additionally, by generating realistic comparative patterns, the GAN also enhances the detection of anomalies, allowing the system to identify variations that can indicate abnormalities. The adaptability of the learning module, triggered by reinforcement learning and constantly updates the detection thresholds based on feedback from dynamic changes in the real world, is what makes CAMM-GAN innovative. The model is thus trustworthy and ensures effectiveness in different real-world applications. Experimental performance tests exhibit a competitive detection accuracy rate of 97.35% on different benchmark datasets such as XD-Violence and UCF Crime. An equal recall and precision for sophisticated anomalies is verified with an F1-score of 96.50%. Given the adaptability, scalability, and robustness of the model, the CAMM-GAN framework fills the requirements of next-generation surveillance while extending situation awareness.</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

A context-aware multi-modal generative adversarial network for real-time anomaly detection in video surveillance

  • 期刊论文

  • 发表日期：

    2025年12月16日

查看全文
  • Pravinth Raja,
  • Dhanalakshmi B K,
  • Rajan T,
  • Vinayakumar Ravi,
  • Norah Saleh Alghamdi

摘要

Anomaly detection by video surveillance needs to ensure public safety in metropolitan and sensitive environments. Most solutions today, though, are saddled with accuracy, scalability, and reliability issues due to the use of fixed detection patterns and single-modal information. In a bid to enhance anomaly detection efficiency in real-time, this paper presents Conditional Attention Mechanism Module (CAMM)- Generative Adversarial Network (GAN) (CAMM-GAN), the first real-time context-aware multimodal GAN framework that (i) conditions attention across video, audio and environmental sensors via a novel CAMM, (ii) uses GAN-based generative regularization to model the long-tailed ‘normal’ manifold, and (iii) incorporates reinforcement learning to adapt thresholds and attention online and UCF-Crime while remaining practical for deployment after standard inference optimizations. CAMM-GAN’s hierarchically combines visual, auditory, and environmental sensor data by eliminating redundant features and promoting context awareness. Temporal activity is abstracted in the shape of 3D Convolutional Neural Networks from video, sequential patterns in audio are abstracted by Recurrent Neural Networks, and decision tree models are used for environmental sensor data such as temperature and smoke. These conflicting data sources are combined into a single framework for multi-modal feature integration through the weighted normalization technique. Additionally, by generating realistic comparative patterns, the GAN also enhances the detection of anomalies, allowing the system to identify variations that can indicate abnormalities. The adaptability of the learning module, triggered by reinforcement learning and constantly updates the detection thresholds based on feedback from dynamic changes in the real world, is what makes CAMM-GAN innovative. The model is thus trustworthy and ensures effectiveness in different real-world applications. Experimental performance tests exhibit a competitive detection accuracy rate of 97.35% on different benchmark datasets such as XD-Violence and UCF Crime. An equal recall and precision for sophisticated anomalies is verified with an F1-score of 96.50%. Given the adaptability, scalability, and robustness of the model, the CAMM-GAN framework fills the requirements of next-generation surveillance while extending situation awareness.