Real-time video violence detection using shallow convolutional neural networks

Detecting violent content in real-time videos has become vital when digital content management and safety for people are of the maximum priority. The research proposes a novel and effective hybrid framework, called Shallow Lightweight Convolutional Attention-based Unified Temporal Network (SLCAUT-Net), proposed to accurately identify violent actions in video streams while sustaining real-time processing competences. To extract spatial features, SLCAUT-Net uses a shallow convolutional neural network (SCNN) backbone, combined with motion-based features derived from optical flow and improved through a temporal attention mechanism. The proposed architecture uses a minimal number of convolutional layers to ensure fast inference while simultaneously capturing temporal dependencies and motion patterns that are critical to distinguishing violent from non-violent behavior. The dataset from Kaggle can be used to support training and valuation by using video clips of real violent and non-violent scenarios. Frame differencing, data augmentation, and lightweight attention modules are employed to increase robustness and reduce overfitting. Experimental assessments validate that SLCAUT-Net attains competitive accuracy ( 97% ) while operating proficiently on low-resource devices, with atypical latency. The addition of temporal attention and motion cues with a shallow framework offers a novel solution to the challenges of violence detection in conditions that are dynamic and unrestricted. The research highlights the potential of hybrid shallow networks in real-time video surveillance and safety-critical applications.