Video Anomaly Detection Based on Frame Prediction of Generative Adversarial Network
Bin Zhao, Boyu Zhao, Pengfei Li · 2021 IEEE International Conference on Emergency Science and Information Technology (ICESIT) · 2021
With the development of society, the application of abnormal behavior detection in the field of public safety has become more and more extensive. We propose a frame prediction video behavior anomaly detection model based on Generative Adversarial Network (GAN). We use the U-net network with the feature storage module and variance attention mechanism as the generator, which not only increases the network's sensitivity to the movement part of the sample, but also reduces the network's learning ability and limits the network's ability to predict abnormal samples. For the discriminant model, we have added a channel and spatial attention mechanism to the Markov discriminator to improve the discrimination ability, which is conducive to improving the quality of future frame generation. Compared with the existing abnormal behavior detection methods, our proposed model achieves excellent detection performance.