Violence Detection: A Serious-Gaming Approach

Internet-of-Things (IoT) technology such as Surveillance cameras are becoming a widespread feature of citizens' life. At the same time, the fear of crime in public spaces (e.g., terrorism) is ever-present and increasing but currently only a small number of studies researched automatic recognition of criminal incidents featuring artificial intelligence (AI), e.g., based on deep learning and computer vision. This is due to the fact that little to none real data is available due to legal and privacy regulations. Consequently, it is not possible to train and test deep learning models. A solution to such shortcoming of datasets is through the use of generative technology and virtual gaming data. Virtual games are a compelling source of data since they can simulate many different scenarios for diverse criminal activities e.g., think of the Grand-Theft Auto (GTA) gaming platform and its opportunities. However, it is not clear whether the synthetically generated data has enough resemblance to the real-world videos to improve the performance of deep learning models in practice. The aim of this work is to investigate the possibilities to identify criminal scenarios with a deep learning model based on video gaming data.

We propose a deep learning violence detection framework using virtual gaming data. The proposed framework is based on a 3-stage end-to-end framework that can be used in crime detection systems. The deep learning framework is divided into two parts: (1) person identification and (2) violence activity recognition. In addition, we introduce a new dataset that allows supervised training of deep learning network models. First, we examine whether the virtual persons were similar enough to persons in the real world.  Second, we examine to what extent video gaming data can be used to identify violent scenarios in the real world. Our results show that virtual persons are just as realistic as persons in the real world. Moreover, our research shows how a serious-gaming approach can be used to identify violent scenarios with an average accuracy 15\% higher than 3 well-known datasets from real-world scenario.