Violence detection in railway environment with modern deep learning approaches and small dataset

Passengers’ security is a continual challenge for public transport operators. For a few years now, cameras are installed inside vehi- cles (train, metro, tramway, bus. . .) but encountered numerous issues. The aim of this study is to propose new video based violence detection algorithms applied to a railway environment exploiting modern deep learning architectures. We have done a domain adaptation of a state-of-the-art architecture selected from the research community on a recorded railway dataset. Then to improve the selected architecture we have proposed to consider a long-term sequence with a recurrent mechanism. Trained architectures have given promising results on selected datasets of the community. The results obtained on our new railway dataset show that 3D convolutional network combined with recurrent networks can model and detect violence in complex railway environment.