Scholarlink

<p>Detecting abnormal behavior in crowded scenes is a critical yet challenging task in intelligent video surveillance due to complex motion patterns, occlusions, and dynamic interactions. To address these challenges, this paper proposes a lightweight, real-time motion-based framework for anomaly detection that does not require data-intensive training. The proposed approach utilizes dense optical flow estimation based on the Farneback algorithm, followed by temporal filtering, adaptive pixel-level motion segmentation, and a novel spatial energy distribution model to enhance anomaly localization. Additionally, an adaptive thresholding mechanism and region-level majority voting are employed to improve detection robustness and reduce false alarms. The framework was evaluated on two benchmark datasets: the UMN dataset and the HAJJv2 dataset, which represent controlled and real-world crowded environments, respectively. Experimental results demonstrate that the proposed method achieves 99.4% accuracy and an mAP of 0.96 on the UMN dataset, and 90.13% accuracy with an mAP of 0.93 on the HAJJv2 dataset, outperforming several state-of-the-art approaches while maintaining real-time performance (28 FPS). These results confirm that the proposed framework provides an effective, interpretable, and computationally efficient solution for anomaly detection in crowded scenes, making it well-suited for deployment in safety-critical 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

A motion-based framework for real-time anomaly detection in crowded scenes

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Doaa Mabrouk,
  • Manal A. Abdel-Fattah,
  • Ahmed Taha

摘要

Detecting abnormal behavior in crowded scenes is a critical yet challenging task in intelligent video surveillance due to complex motion patterns, occlusions, and dynamic interactions. To address these challenges, this paper proposes a lightweight, real-time motion-based framework for anomaly detection that does not require data-intensive training. The proposed approach utilizes dense optical flow estimation based on the Farneback algorithm, followed by temporal filtering, adaptive pixel-level motion segmentation, and a novel spatial energy distribution model to enhance anomaly localization. Additionally, an adaptive thresholding mechanism and region-level majority voting are employed to improve detection robustness and reduce false alarms. The framework was evaluated on two benchmark datasets: the UMN dataset and the HAJJv2 dataset, which represent controlled and real-world crowded environments, respectively. Experimental results demonstrate that the proposed method achieves 99.4% accuracy and an mAP of 0.96 on the UMN dataset, and 90.13% accuracy with an mAP of 0.93 on the HAJJv2 dataset, outperforming several state-of-the-art approaches while maintaining real-time performance (28 FPS). These results confirm that the proposed framework provides an effective, interpretable, and computationally efficient solution for anomaly detection in crowded scenes, making it well-suited for deployment in safety-critical surveillance applications.