PREDICTING ROBBERY BEFORE IT HAPPENS: AN AI-DRIVEN APPROACH FOR INDOOR SURVEILLANCE

Abstract

For video surveillance systems to avoid incidents and safeguard assets, crime prediction is necessary. In this regard, our paper suggests the first artificial intelligence method for predicting and detecting Robbery Behaviour Potential (RBP) in an interior camera. Three detection modules—head cover, crowd, and loitering detection modules—are the foundation of our approach, which enables prompt response and deters robberies. The YOLOV5 model is retrained using the manually annotated dataset we collected to create the first two modules. Furthermore, we provide a new definition for the DeepSORT algorithm-based loitering detection module. After converting expert information into rules, a fuzzy inference system makes a final determination about the likelihood of robbery. The robber's various style, the security camera's variable viewpoint, and the poor quality of the video photos make this tedious. We successfully completed our experiment using actual video surveillance footage, achieving an F1-score of 0.537. Therefore, we design a threshold value for RBP to assess video pictures as a robbery detection issue in order to do an experimental comparison with the other relevant research. Assuming this, the experimental findings clearly reflect an F1-score of 0.607, indicating that the suggested approach performs much better in identifying the heist than the robbery detection methods. We firmly think that by anticipating and averting robbery incidents, the implementation of the suggested strategy might reduce the harm caused by robberies in a security camera control centre. However, the human operator's situational awareness improves, and additional cameras may be controlled.