In this paper, we propose FGITNet, an innovative Two-Stream Gated Network designed for accurate violence detection in surveillance videos. Leveraging both Inflated 3D ConvNet and Transformer architectures, FGITNet extracts spatiotemporal features by processing RGB and Optical Flow frames simultaneously. A key contribution of our model is a novel background suppression mechanism that effectively filters out non-motion elements, enhancing motion feature detection while maintaining computational efficiency. This two-stream approach achieves state-of-the-art performance on the RWF-2000 and Hockey datasets, surpassing existing methods by 4% on RWF-2000, reaching an accuracy of 93.75%. Our method demonstrates significant potential for real-time applications in intelligent security systems, contributing to the field of design and intelligent engineering.