End-to-End Video Violence Detection with Transformer

Human action recognition is a widely investigated field in computer vision. Violence automatic detection is a subset of action recognition, which deserves special attention because of its wide applicability in unmanned security monitoring systems. This paper presents an end-to-end model, which introduces Transformer for human pose estimation and 3d convolutional neural network to capture motion present in spatial-temporal dimension. We train a 3d convolutional neural network to learn spatial-temporal features of human keypoint sequences which are the outputs of Transformer block. Our proposed model achieves an accuracy of 89% on the test set of large-scale video database RWF-2000, and obtains an accuracy of 93% on our own school violent video database. Our experiment result shows that transformer-based approach can be used in video violence detection.