Deep Context-Aware Feature Extraction for Anomaly Detection in Surveillance Videos

Surveillance video analysis plays a crucial role in ensuring public safety and security. Developing a context-aware framework for anomaly detection in surveillance videos is motivated by the need for enhanced security, safety, and efficiency in various domains. Context-aware anomaly detection depends on spatiotemporal features that help the model understand the context of anomalies in surveillance videos. This study aimed to provide a novel deep learning-based context-aware approach to feature extraction to detect anomalies in surveillance videos. The proposed method integrates ResNet50 for spatial feature extraction and 3D Convolutional Neural Network (CNN) for temporal feature extraction. This method identifies six anomalous activities, namely abuse, arrest, fighting, robbery, shooting, and road accidents, using the UCF-Crime dataset. The proposed integrated ResNet50 and 3D CNN model achieves promising accuracy for the six classes, such as 95% for abuse, 93% for arrest, 95.22% for fighting, 94.44% for robbery, 93% for shooting, and 94.22% for road accidents. By combining spatiotemporal features, the proposed model detects anomalies in behavior and unexpected movements, which makes it useful for security monitoring where deviations from normal behavior indicate anomalous events. This research contributes to advancing the capabilities of surveillance systems, enhancing public safety, and enabling proactive security measures in diverse urban environments.