A Frame-Based Feature Model for Violence Detection from Surveillance Cameras Using ConvLSTM Network

Abstract

Vision-based action detection is a puzzling research topic of computer vision and pattern recognition. Violence detection in video has found a particularly fundamental application in real-life scenes, aiming at monitoring and protecting the safety of people and pedestrians. In previous years, violence detection in surveillance cameras gained more and more attention from researchers, due to a huge amount of video data collected by Closed-circuit television (CCTV) cameras that covers different places such as schools, hospitals, shopping malls, streets, etc… However, the manual control and detection of violent behaviors and report of those incidents on time based on videos is very tiresome and laborious for operators which may results in the loss of lives. With the rise of deep learning and its applications in different domains of human daily lives, automatic detection models for abnormal behaviors or violence incidents are highly needed for this concern. In this paper, we address this research problem and explore ConvLSTM network as a solution. We conducted extensive experiments on six benchmark datasets, where our method shows superior performance on all these datasets by improving the state-of-the-art performance accuracy for violence detection.