Real Life Violence Detection in Surveillance Videos using Spatiotemporal Features

Automatic violence detection has remarkable importance from practical and academic point of view. Generally speaking, detecting violence in a crowded locality, via computational approaches, is challenging owing to rapid movements, overlapping characteristics, obstructed scenery, and scattered backgrounds. Fortunately, Deep Learning techniques can detect anomalies to a certain extent. Furthermore, their popularity, as a paradigm to detect violence, is growing at a tremendous pace. The aim of such approaches is to develop a method that recognizes violence and evokes an alarm so that immediate assistance can be provided. This paper is aong the same line of thought. This article presents a Convolution Neural Network (CNN) and Recurrent Neural Network (RNN) based approach for violence detection by learning the detailed features in videos. The spatio-temporal features extracted from the combination of InceptonV3 pre-trained model and late LSTM architecture yielded a 97.5% accuracy thereby, proving its superiority over existing methods in literature.