Abstract:Crowd management and surveillance have emerged as critical challenges in public safety, especially during large gatherings, protests, and events. Traditional manual surveillance methods are inefficient, error-prone, and slow to respond to dynamic crowd behaviours. In this research, we propose a real-time, automated crowd behaviour analysis and alert system leveraging deep learning models and the Flask web framework. Our system integrates a custom Convolutional Neural Network (CNN), a fine-tuned VGG16 model, YOLOv8n classification, and YOLOv8 object detection for behaviour recognition and headcount estimation. The Flask application serves as the front-end, facilitating video upload, webcam live streaming, and visualization of results. The system automatically triggers alarms and email notifications upon detection of violent activities. Experimental evaluation demonstrates a classification accuracy of 99.23% using VGG16 and near real-time inference at 30 FPS with YOLOv8n. This work establishes a foundation for deploying AI-driven surveillance systems capable of reducing manual effort, enhancing situational awareness, and ensuring public safety in crowded environments.

Keywords: Crowd Behaviour, Deep Learning, YOLOv8, Flask Web Framework, Real-Time Surveillance, Public Safety, Convolution neural Network (CNN), VGG16.

| DOI: 10.17148/IARJSET.2025.12583