Adaptive and effective spatio-temporal modelling for offensive video classification using deep neural network

Security cameras have recently been widely implemented in public locations, and the overall crime rate has decreased dramatically due to these omnipresent gadgets. These cameras are typically employed to offer cues and evidence after crimes have occurred rather than to prevent or deter criminal activity in the first place. Manually monitoring a lot of video footage from surveillance cameras takes significant time and effort. As a result, it is critical to automatically recognise aggressive actions from video signals. It is also critical to detect violence in videos to protect children from inappropriate content. This paper tackles the difficult subject of detecting violence in videos. Unlike previous work focusing on merging multimodal features, we take it further by including and utilising visual subtypes connected to violence. The proposed approach is implemented using Jupyter notebook and Tensorflow, with better accuracy of 76.79% on the proposed database test set.