Deep Learning-Based Violence Detection from Videos

Violence has been one of the major concerns among human interactions. Violent activities turn out to be worse in public places like parks, halls, stadiums, and many more. The presence of efficient detection algorithms is the need of the hour as unusual events such as fights have been comparatively studied less. The existing system requires manual monitoring of videos from surveillance cameras. Recognition of violent interactions is important to develop automated video monitoring systems. Approaches based on deep Learning promise better results in recognition of images and human actions. Our proposed deep learning-based model uses CNN and LSTM based on DarkNet19 architecture as the pretrained model. The proposed model achieves 95 and 100% accuracy on the benchmark hockey and Peliculas datasets. Unlike most previous works which involve the use of single-frame models (models which do not include temporal features), our model is able to learn temporal features as well as spatial features. The two datasets, namely Hockey Fight and Movie datasets, contain motions of sudden camera recordings and are challenging datasets to work on. It is also observed that different variations in optical flow were taken into consideration for the work in hand. However, the existing system was unable to utilize an efficient pretrained model for the task of violence detection. The proposed approach puts forward a more efficient pretrained model architecture which is less studied, but is very effective than traditional methods.