Real-Time Violence Detection in Video Footage Using a Mobile-Friendly CNN-Based Model

Detecting violence in video content, particularly within domestic environments, presents an ongoing challenge in both social and technological contexts. This paper proposes a lightweight deep learning framework for real-time violence detection, optimized for mobile and edge deployment. The approach is based on MoViNet-A0, evaluated in both Base and Stream configurations, and is complemented by a custom Conv2D-based baseline designed for ultra-low-latency inference. All models were trained and validated on the AIRTLab dataset, which includes 350 annotated videos representing violent and non-violent scenes. The Mo YiN et-A0 Base model achieved a validation accuracy of 92.8%, while the Conv2D-based model reached 89.6% validation accuracy, along with a precision and F1-score close to 90%. Performance benchmarks conducted on Android devices and desktop platforms show that real-time inference is feasible, with latencies as low as 0.9 seconds per 10-frame sequence on mid-range smartphones. The entire pipeline has been designed for mobile deployment, and integration into a functional prototype application is currently in progress, aiming to enable real-time violence detection directly on mobile devices.