This paper introduces an automated method for detecting violence. By employing cutting-edge deep learning algorithms and real-time video surveillance analytics, it improves public safety in smart cities. The proposed method integrates LSTM networks with CNNs to gather geographical and temporal data from surveillance footage in order to detect violent scenarios, including rioting, physical assaults, and unexpected crowd hostility. In order to distinguish between violent occurrences and arbitrary human interactions in a variety of illumination, occlusion, and crowd density scenarios, the model employs attention techniques for feature representation and transfer learning with pre-trained architectures. In experiments conducted on benchmark datasets, accuracy, precision, and recall are maintained, while computing efficiency is preserved. Consequently, it is the optimal choice for smart city infrastructure that is connected to the periphery. This research improves IUSS by creating a proactive, scalable, and automated method for the detection of violent incidents, which enables quicker reactions and enhanced crime prevention.