Real-Time Anomaly Identification in Surveillance Videos Using Object Tracking and Spatio-Temporal Graph Learning

Real-time anomaly identification in surveillance recordings is essential for intelligent monitoring systems' public safety and proactive event response. Most video anomaly detection techniques use frame-level representations or constructed motion characteristics, which cannot describe complicated item interactions and long-term temporal correlations in crowded scenes. This research offers an object tracking and spatio-temporal graph learning-based real-time anomaly detection framework to solve these restrictions. First, the suggested system recognizes and tracks many objects in real time to retain object identities between frames. Each tracked object's discriminative appearance, motion, and spatial attributes are combined to create a dynamic spatio-temporal graph that simulates inter-object interactions and temporal evolution. A Spatio-Temporal Graph Neural Network (ST-GNN) learns normal behavior by transmitting spatial and temporal messages. Deviations from learnt normal behavior embeddings are used to calculate anomaly scores, which accurately identify aberrant events. Experimental evaluation on the UCF-Crime dataset shows that the proposed framework outperforms frame-level, sequence-based, and graph-based anomaly detection methods with an AUC of 92.8%, accuracy of 91.2%, and F1-score of 90.9%. The results show that precisely describing object interactions and temporal dynamics improves detection. The suggested real-time system is robust to environmental changes, making it suitable for intelligent surveillance applications