Real-time Aggressive Action Recognition using Deep Learning in a Multifarious Setting

The epidemic of violent crimes worldwide necessitates an active-based video surveillance network to combat these criminal acts. In this context, autonomously detecting aggressors and weapons is crucial in modelling human-weapon interactions for aggressive action recognition. However, current object detectors and human-object interaction detectors using deep learning cannot reliably capture surveillance-based humans and weapons in multifarious and complex scenarios. To address these problems, this research puts forward a novel surveillance-based human-object interaction (HOI) detector with an aggressive HOI dataset for accurate aggressive action detection. In the end, the outcomes of this research could potentially pioneer real-time crime detection in video surveillance.