Resource Constrained, Fast Convergence Training for Violence Detection in Video Streams

This paper addresses the automated identification of violent acts from CCTV video streams using a Deep Learning model under constrained resources. While this process typically involves a powerful setup, it is useful to accelerate the training and get accurate results using more modest computational resources that would bring automatic recognition of violent acts closer to common surveillance resources. Our results provide 94.98% accuracy, on par with the state-of-the-art, but at a fraction of the training time. This translates into lower energy requirements and allows a broader deployment on large scale (urban) autonomous surveillance networks while providing an increased privacy towards citizens and lower chances of abuse from authorities.