Detection of Violent Activities Using Deep Learning Algorithms

Abstract

The creation of a method for violence detection in surveillance footage using automatic analysis is crucial. In this study, we propose a deep neural network to recognize violent videos. A convolutional neural network and an ImageNet model that has already been trained are used to extract frame level characteristics from a movie. Then, using a long short-term memory variation that makes use of fully connected layers and leaky rectified linear units, the frame level features are aggregated. Convolutional neural networks are capable of recording localized spatio-temporal information that allow the analysis of local motion in the video, in addition to long short-term memory. On three common benchmark datasets, the accuracy of recognition is used to further assess the performance. We also contrasted the findings of our system with those from other methodologies to ascertain the capabilities of our proposed model. The suggested solution outperforms cutting-edge techniques while processing the videos in real-time.