aspg-28-3858-1792156266 20261016131106 American Scientific Publishing Group admin@americaspg.com American Scientific Publishing Group Journal of Artificial Intelligence and Metaheuristics JAIM 2833-5597 2025 9 2 Muhammad Ahsan School of Mathematical Sciences, Jiangsu University, Jiangsu 212013, China The rapid development of smart-city infrastructure and the complexity of urban environments increase the need for automated systems that detect violence in surveillance footage. This paper presents a deep-learning architecture for detecting violence and weapon-related activity in CCTV footage using the Smart-City CCTV Violence Detection (SCVD) dataset. MobileNetV2 extracts spatial features from video frames through TimeDistributed layers, and a stacked Long Short-Term Memory (LSTM) network models temporal dependencies. The system processes 15-frame sequences at a resolution of 128 × 128 pixels. Batch normalization and dropout are used to improve generalization and limit overfitting, while a sigmoid output produces binary classifications. On the SCVD evaluation, the model achieved 99.58% accuracy and a binary cross-entropy loss of 0.0139. The Normal class achieved precision, recall, and F1-score values of 0.99; the Violence class metrics were reported as 1.00 after rounding. These results indicate the potential of compact spatiotemporal models for reliable surveillance and deployment on edge devices such as smart cameras and embedded systems. 2025 14 23 3858 https://creativecommons.org/licenses/by/4.0/ 10.54216/JAIM.090202 https://www.americaspg.com/journal/28/article/3858 https://www.americaspg.com/storage/articles/manuscripts/01M4Y5M57P2FJ6BN2E7FX66AZH.pdf