Scholarlink

<p>This paper presents a novel data stream-based approach for video anomaly detection in surveillance scenarios, addressing the limitations of traditional deep learning and batch processing methods. Recognizing surveillance video's continuous and evolving nature, the proposed method leverages a data stream methodology, processing incoming video segments incrementally. A key contribution lies in a weakly supervised temporal feature extraction process. This process employs Discrete Cosine Transform (DCT) to project video segments into the frequency domain. Then, it utilizes Accelerated Particle Swarm Optimization (APSO) for feature selection, effectively reducing dimensionality while preserving crucial information and enhancing computational efficiency. The selected features are then fed into an ensemble of Unsupervised Feature Representative Online Sequential Extreme Learning Machines (UFROS-ELMs), which adapt to evolving data distributions through an integrated concept drift detection and adaptation mechanism. This dynamic adaptation is critical for handling the non-stationary nature of real-world surveillance data. Experimental results on benchmark datasets, UCF-Crime and UCSD Pedestrian, demonstrate the proposed method's superior performance compared to state-of-the-art approaches, achieving higher accuracy and lower error rates while maintaining computational efficiency suitable for real-time applications. This streaming data-based approach achieves cost-effective and more effective anomaly detection by synergistically combining DCT/APSO feature extraction with adaptive UFROS-ELM learning.</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 novel streaming data-based video anomaly detection method with adaptive feature selection

  • 期刊论文

  • 发表日期：

    2025年12月21日

查看全文
  • Ozge Aydogdu,
  • Mehmet Cemil Aydogdu

摘要

This paper presents a novel data stream-based approach for video anomaly detection in surveillance scenarios, addressing the limitations of traditional deep learning and batch processing methods. Recognizing surveillance video's continuous and evolving nature, the proposed method leverages a data stream methodology, processing incoming video segments incrementally. A key contribution lies in a weakly supervised temporal feature extraction process. This process employs Discrete Cosine Transform (DCT) to project video segments into the frequency domain. Then, it utilizes Accelerated Particle Swarm Optimization (APSO) for feature selection, effectively reducing dimensionality while preserving crucial information and enhancing computational efficiency. The selected features are then fed into an ensemble of Unsupervised Feature Representative Online Sequential Extreme Learning Machines (UFROS-ELMs), which adapt to evolving data distributions through an integrated concept drift detection and adaptation mechanism. This dynamic adaptation is critical for handling the non-stationary nature of real-world surveillance data. Experimental results on benchmark datasets, UCF-Crime and UCSD Pedestrian, demonstrate the proposed method's superior performance compared to state-of-the-art approaches, achieving higher accuracy and lower error rates while maintaining computational efficiency suitable for real-time applications. This streaming data-based approach achieves cost-effective and more effective anomaly detection by synergistically combining DCT/APSO feature extraction with adaptive UFROS-ELM learning.