Crowd Density Mapping and Anomaly Detection using YOLOv8 and DEEPSORT

Crowd Density Mapping and Anomaly Detection using YOLOv8 and DEEPSORT

Dr. Deepali Ujlambkar1*,  Gaurav Shirke2, Suraj Pawar3, Tejas Pawar4, and Raviraj Shingate5

1Assistant Professor, Department of Computer Engineering, AISSMS College of Engineering, Pune, India.

2,3,4,5 Student, Department of Computer Engineering, AISSMS College of Engineering, Pune, India.

{gauravshirke895, surajpawar0216, tejasanandapawar2003, ravirajshingate173}@gmail.com

Abstract:
Effective crowd management in high-density public spaces remains a critical challenge, especially during large-scale events or emergencies. This study introduces a real-time system that integrates crowd density estimation and behavioral anomaly detection using deep learning and video surveillance. The proposed framework leverages YOLOv8 for high-precision person detection and DeepSORT for continuous multi-object tracking. Anomalies—such as panic movements, physical altercations, and prolonged immobility—are identified using motion trajectory analysis and rule-based behavioral classification. Crowd density is categorized into low, medium, or high based on per-frame detections, and visualized through a dynamic Google Maps interface. The system also issues immediate alerts for critical events using SMTP-based notifications. Evaluated on benchmark datasets including ShanghaiTech, UCF-QNRF, and a custom-labeled Roboflow dataset, the model achieves a detection precision of 86.2% and supports near real-time processing with optimized latency on both CPU and GPU platforms. This approach provides scalable, location-aware crowd analytics, making it highly applicable for smart surveillance, urban safety, and emergency response systems.

Keywords: Crowd density estimation, Anomaly detection, YOLOv8, LSTM, CNN (Convolutional Neural Network), Real-time monitoring, Crowd detection, Interactive heatmaps, Public safety, CCTV surveillance.