The increasing need for safety in public and private spaces has led to the development of advanced surveillance systems capable of detecting violent incidents in real time. As violent events can escalate quickly, it is critical for trained professionals to intervene at the earliest detection point. Traditional surveillance systems rely on basic video monitoring, but advancements in machine learning and deep learning models have significantly enhanced the capabilities of these systems. This study focuses on evaluating four deep learning models—YOLOv8, MobileNetV3, and Transformer—using publicly available video frames to detect violent actions. The models were trained with optimized parameters, and their performance was rigorously assessed based on accuracy and detection capability. The Transformer model was found to outperform the other models, achieving the highest accuracy in identifying violent incidents. This suggests that Transformer-based models can provide superior performance for real-time violence detection in dynamic video environments. YOLOv8 and MobileNetV3, while effective, did not achieve the same level of accuracy, highlighting the potential of Transformer models for future surveillance applications. These findings emphasize the importance of selecting the right model for effective violence detection and contribute to the ongoing evolution of AI-powered surveillance systems that can detect and respond to violent situations swiftly and accurately.

Keywords: Violence detection, video surveillance, deep learning, YOLOv8, MobileNetV3, Transformer, machine learning, Non-Violence detection.