Lightweight MobileNetV4-ConvLSTM Architecture for Spatiotemporal Video Violence Detection

Research Context: Violence remains a global challenge, impacting millions annually and requiring scalable technological solutions for prevention and detection. Practical Problem: Manual analysis of surveillance video is labor-intensive and prone to error, hindering timely identification of violent incidents. Developing models that combine high accuracy with computational efficiency is essential for practical deployment. Proposed Solution: This work proposes a hybrid MobileNetV4–ConvLSTM model that extracts spatial and temporal features from video frames, providing a lightweight yet robust architecture for automatic violence detection. Related IS Theory: This study is grounded in Task–Technology Fit and Socio-Technical theories, highlighting the alignment between the task of video-based violence detection and the technological sup- port needed, integrating people, processes, and technology to enhance public-safety decision-making. Research Method: This applied experimental study trained and evaluated the proposed model on the RWF-2000 dataset. Performance was assessed primarily through classification accuracy, while computational efficiency was measured using experiments conducted on AWS EC2 in- stances. Summary of Results: The proposed model achieved 96.94% accuracy with approximately 5.6 million parameters, demonstrating strong performance with reduced computational cost. Contributions and Impact in the IS Area: The study advances research in video-based violence detection by presenting a compact, high-performing architecture suitable for real-time surveillance and public safety applications.