Scholarlink

<p>Crowding disasters are prone to happen in mass public events such as the Hajj pilgrimage, usually preceded by some counter-flow crowd behavior. Current crowd management systems are mostly reactive based on density measurement or anomaly detection only after dangerous crowd behavior has occurred, having insufficient time for intervention. The objectives of the current research are to design a proactive system that has the ability to forecast congestions by analyzing directional flows in real-time. The current research has proposed a pilot-scale validation of the proposed method using a dataset that has been specially designed for the HAJJv2 video corpus, in which pilgrims are classified according to whether they are in a ‘normal direction’ or in the ‘opposite direction. We created a tailored dataset out of the HAJJv2 video dataset and annotated pilgrims into two categories: 'normal direction' and 'opposite direction'. The technique uses a complex preprocessing pipeline with Non-Local Means Denoising and Contrast Limited Adaptive Histogram Equalization (CLAHE) to preprocess the images for processing under harsh Hajj conditions. We trained and tested four top-of-the-line You Only Look Once (YOLO) architectures (YOLOv8n, YOLOv9c, YOLOv10n, and YOLO11n) with rigorous data augmentation to tackle severe class imbalance and enhance model generalization. The models performed exceptionally well under extreme crowd situations, heavy occlusions, and a severe 14.3:1 class imbalance. YOLO11n emerged as the best-performing model on the unseen test set, achieving the highest Mean Average Precision (mAP@50) of 0.935. Meanwhile, YOLOv8n presented the most balanced operational profile, delivering a highly competitive test mAP@50 of 0.932 while requiring the lowest training time (1.16 h), making it the optimal choice for rapid iteration and resource-constrained research. Statistical validation via the Friedman test confirmed that the performance variance across the models was statistically significant (p = 0.0140). Crucially, all models successfully detected the challenging minority 'Normal' class, validating the effectiveness of our class-balancing countermeasures. The study can show the potential of deploying deep learning in directional crowd flow analysis as a predictive indicator of crowd safety. From the reactive monitoring model to an early warning system with proactive counter-flow identification, the introduced framework proposes a proactive mechanism that detects early indicators of directional anomalies. By identifying counter-flow patterns, the system offers the theoretical potential to provide authorities with a critical intervention window, estimated at 2–5 min based on established crowd dynamics literature, before dangerous density thresholds are reached. It provides crowd accident avoidance with actionable intelligence and significantly enhances pilgrim safety at the Hajj and other mass gatherings.</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

AI-powered directional anomaly detection in dense hajj crowds for proactive safety management

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月2日

查看全文
  • Ryan Alturki,
  • Mohammad Wedyan,
  • Ahmad Nasayreh

摘要

Crowding disasters are prone to happen in mass public events such as the Hajj pilgrimage, usually preceded by some counter-flow crowd behavior. Current crowd management systems are mostly reactive based on density measurement or anomaly detection only after dangerous crowd behavior has occurred, having insufficient time for intervention. The objectives of the current research are to design a proactive system that has the ability to forecast congestions by analyzing directional flows in real-time. The current research has proposed a pilot-scale validation of the proposed method using a dataset that has been specially designed for the HAJJv2 video corpus, in which pilgrims are classified according to whether they are in a ‘normal direction’ or in the ‘opposite direction. We created a tailored dataset out of the HAJJv2 video dataset and annotated pilgrims into two categories: 'normal direction' and 'opposite direction'. The technique uses a complex preprocessing pipeline with Non-Local Means Denoising and Contrast Limited Adaptive Histogram Equalization (CLAHE) to preprocess the images for processing under harsh Hajj conditions. We trained and tested four top-of-the-line You Only Look Once (YOLO) architectures (YOLOv8n, YOLOv9c, YOLOv10n, and YOLO11n) with rigorous data augmentation to tackle severe class imbalance and enhance model generalization. The models performed exceptionally well under extreme crowd situations, heavy occlusions, and a severe 14.3:1 class imbalance. YOLO11n emerged as the best-performing model on the unseen test set, achieving the highest Mean Average Precision (mAP@50) of 0.935. Meanwhile, YOLOv8n presented the most balanced operational profile, delivering a highly competitive test mAP@50 of 0.932 while requiring the lowest training time (1.16 h), making it the optimal choice for rapid iteration and resource-constrained research. Statistical validation via the Friedman test confirmed that the performance variance across the models was statistically significant (p = 0.0140). Crucially, all models successfully detected the challenging minority 'Normal' class, validating the effectiveness of our class-balancing countermeasures. The study can show the potential of deploying deep learning in directional crowd flow analysis as a predictive indicator of crowd safety. From the reactive monitoring model to an early warning system with proactive counter-flow identification, the introduced framework proposes a proactive mechanism that detects early indicators of directional anomalies. By identifying counter-flow patterns, the system offers the theoretical potential to provide authorities with a critical intervention window, estimated at 2–5 min based on established crowd dynamics literature, before dangerous density thresholds are reached. It provides crowd accident avoidance with actionable intelligence and significantly enhances pilgrim safety at the Hajj and other mass gatherings.