Detection of violent activity in surveillance system using different deep learning techniques

Abstract

In the past decade, surveillance cameras have been a necessary integration for security measures in all types of localities. The omnipresence of these devices has substantially aided in tackling violent criminal activities. In larger systems, continuous manual monitoring becomes a cumbersome task and often causes delayed response. Therefore, automated recognition of aggressive activities in surveillance systems can enhance the remote monitoring experience and increase the preciseness of response. Previous experiments on various deep-learning techniques and Convolutional Neural Networks (CNN) have tackled the challenge by identifying potential violent activities in real-time with good accuracy. The aim of this research is to benefit from reduced computational cost while maintaining optimality for practical implementation in real life. Hence, in this study, preliminarily a lightweight yet highly effective CNN model has been proposed that extracts spatial features by 2D convolutions. Later on several custom models based on combinations of CNN and RNN architectures have been developed for spatio-temporal features from the videos. The models have undergone robust tuning and training and are capable of accurately extracting frame-level and temporal-level features based on the architectural types. They have been then conclusively evaluated on a combination of multiple benchmark datasets to compare how well each of them performs. In conclusion, the proposed spatial feature-based model obtained an outstanding test accuracy of 99.6% and the best spatio-temporal feature-based model in terms of performance attained a test accuracy of 98.75%.