Abstract
This study developed an automatic video-based violence detection system by combining the MobileNetV2 architecture and Bidirectional Long Short-Term Memory (BiLSTM). MobileNetV2 is used to efficiently extract spatial features from each video frame, while BiLSTM is utilized to understand the temporal relationships between frames so that violence patterns can be recognized more accurately. The model is designed to classify videos into two classes: “Violence” and “Non-Violence.” The training process showed a consistent increase in accuracy with a decrease in loss values, without any indication of overfitting. The resulting model achieved an accuracy of 92%, with high precision and recall values, especially in detecting violent actions. The system is implemented in real-time using input from a live camera. When violence is detected, the system automatically displays a visual warning in the form of a red screen with the text “VIOLENCE DETECTED!” and saves a frame clip with a timestamp as documentation. Testing shows that the system can distinguish violent contexts effectively without producing significant false positives. With reliable performance and quick response times, this system has great potential for application in CCTV surveillance in public areas, schools, and conflict-prone regions as an efficient, proactive, and adaptive AI-based solution.