Real-time detection of violent behavior through surveillance technologies is increasingly important for public safety. This study tackles the challenge of automatically distinguishing violent from non-violent activities in continuous video streams. Traditional surveillance depends on human monitoring, which is time-consuming and error-prone, highlighting the need for intelligent systems that detect abnormal behaviors accurately with low computational cost. A key difficulty lies in the ambiguity of defining violent actions and the reliance on large annotated datasets, which are costly to produce. Many existing approaches also demand high computational resources, limiting real-time deployment on resource-constrained devices. To overcome these issues, the present work employs the lightweight MobileNet deep learning architecture for violence detection in surveillance videos. MobileNet is well-suited for embedded devices such as Raspberry Pi and Jetson Nano while maintaining competitive accuracy. In Python-based simulations on the Hockey Fight dataset, MobileNet is compared with AlexNet, VGG-16, and GoogleNet. Results show that MobileNet achieved 96.66% accuracy with a loss of 0.1329, outperforming the other models in both accuracy and efficiency. These findings demonstrate MobileNet’s superior balance of precision, computational cost, and real-time feasibility, offering a robust framework for intelligent surveillance in public safety monitoring, crowd management, and anomaly detection.