The increasing adoption of intelligent video surveil-lance systems in public security settings has intensified the need for automated approaches capable of monitoring crowded envi-ronments effectively. In conventional surveillance infrastructures, human operators are required to supervise multiple video streams simultaneously, which limits the reliable detection of infrequent yet critical events and often results in reduced performance due to fatigue and attention constraints. To mitigate these limitations, a two-level deep learning framework is proposed for the analysis of dense crowd scenes, with the objective of identifying both abnormal crowd behavior and explicit security threats such as weapons. In the first stage, a YOLOv8-based object detection model is utilized for real-time processing to enable person localization, crowd density estimation, and the detection of visible weapons. Frames that contain relevant activity are subsequently processed using an EfficientNet-B0 network, chosen for its favorable balance between recognition accuracy and computa-tional efficiency, to extract spatial feature representations. These features are then analyzed using an attention-based bidirectional gated recurrent unit (BiGRU) network, which models temporal dependencies across frame sequences to classify overall crowd behavior. The incorporation of an attention mechanism allows the system to emphasize temporally informative frames, thereby improving sensitivity to the onset and progression of abnormal activities. The behavioral classification module is trained and evaluated on the UCF-Crime dataset, and the experimental results demonstrate reliable performance in practical anomaly detection scenarios. Furthermore, the complete framework is implemented as a real-time, modular web application using the Flask framework, incorporating an interactive dashboard, alert generation, and event logging. This implementation illustrates the feasibility of deploying the proposed system in real-world surveillance environments.

Index Terms—Video anomaly detection, weapon detection, crowd monitoring, deep learning, intelligent video surveillance, YOLOv8, EfficientNet-B0, attention mechanism, bidirectional GRU, real-time systems.