Using deep learning algorithms to detect violent activities

Abstract

It is of extensive importance to develop a technique for automatic surveillance video analysis to recognize the presence of violence. In this work, to identify violent videos, we put forward a deep neural network. For extracting frame level features from a video, a convolutional neural network is used with a pre-trained ImageNet model. The characteristics of the frame level are then aggregated using a long short-term memory variant that uses fully connected layers and leaky recti ed linear units. Together with the long short-term memory, the convolutional neural network is capable of capturing localized spatio-temporal features that enable the analysis of local motion in the video. The performance is further evaluated in terms of accuracy of recognition on three standard benchmark datasets. In order to determine the capabilities of our proposed model, we also compared our system results with other techniques. The approach proposed outperforms state-of - the-art methods while processing the videos in real time.