Robust crowd anomaly detection via hybrid ensemble learning for real-world surveillance

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Crowd Anomaly Detection (CAD) is vital for public safety in crowded areas.
• Existing methods often struggle with generalizability, high computational demands, and small datasets.

Purpose of the Study:

• To develop a novel hybrid ensemble learning framework for improved Crowd Anomaly Detection.
• To enhance real-time surveillance capabilities in data-constrained environments.

Main Methods:

• Integrated YOLOv7 for real-time crowd detection with Random Forests and Gradient Boosting classifiers.
• Extracted spatial and motion features using optical flow from YOLO-detected regions.
• Optimized the ensemble model using the Adam optimizer for data-constrained scenarios.

Main Results:

• Achieved near-perfect accuracy of 99.89% on the University of Minnesota (UMN) benchmark dataset.
• Demonstrated strong generalizability and robustness on a custom real-world supermarket dataset.
• Established a new standard for small-scale CAD.

Conclusions:

• The proposed hybrid ensemble framework offers a practical solution for intelligent, real-time surveillance.
• The model shows significant improvements in accuracy and robustness for Crowd Anomaly Detection.
• This approach addresses limitations of existing CAD methods, particularly in data-scarce situations.