Detection and Localization of Anomalies in Crowded Environments

Abstract—This project introduces a lightweight, offline system for detecting anomalies in human movement within surveillance video. Implemented in Python, it leverages Farneback dense optical flow to extract motion features specifically magnitude and direction from consecutive frames. A Decision Tree Classifier, trained on labeled motion data, identifies abnormal behaviors within frame segments. Anomalies are visually highlighted in real-time using color-coded overlays, offering intuitive feedback. The system is optimized for resource-limited devices like laptops and Raspberry Pi, making it suitable for edge deployment. Its modular design supports easy model replacement or enhancement. Future improvements may include deep learning integration, better tracking, and more user-friendly interfaces. Index Terms—Anomaly Detection, Optical Flow, Surveillance Video Analysis, Motion Feature Extraction, Decision Tree Classifier, Real-Time Processing, Offline Monitoring, Farneback Algorithm, UCSD Ped1 Dataset, Foreground Segmentation.