Stampede detection and crowd analysis using CNN-LSTM and farneback optical flow.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Public Safety

Background:

• Stampede incidents in crowded environments pose significant public safety risks, leading to casualties and disruptions.
• Existing methods for crowd monitoring often lack the granularity needed for effective early warning systems.

Purpose of the Study:

• To develop and validate a novel, data-driven framework for automated stampede detection and crowd risk classification.
• To provide a more granular assessment of crowd states beyond traditional binary models.

Main Methods:

• Integration of Farneback optical-flow computation with a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) architecture.
• Utilization of combined UCSD Anomaly detection and Agoraset datasets for comprehensive crowd behavior and density analysis.
• Training and evaluation on 10,000 annotated frames with preprocessing and augmentation for robustness.

Main Results:

• Achieved a high accuracy of 99.75% in classifying crowd states into four risk levels: normal, moderate, dense, and risky.
• Demonstrated robustness under domain shift conditions through cross-dataset evaluation on the UMN benchmark.
• The system offers a granular assessment of crowd risk, outperforming traditional binary models.

Conclusions:

• The proposed framework shows strong potential for real-time deployment in public event management and emergency response systems.
• Current limitations include computational latency and challenges in detecting stampedes in ultra-dense, occluded crowd scenarios.
• Further research is needed to address computational efficiency and performance in complex, occluded environments.