Anomaly Detection in Surveillance Videos Based on Two-Stream Inflated 3d Conv Net and Weakly Supervised Learning

Abstract:
A variety of surveillance videos put forward higher requirements for anomaly detection. Among them, the identification of abnormal events in surveillance video has always been a challenging task. In the real world, a variety of factors will cause great difficulties in anomaly recognition. On this basis, we invoke a two-stream inflated 3D convolutional network and a weakly supervised anomaly detection framework called Anomaly Regression Net (AR-Net) to detect surveillance video. The comprehensive experiment is carried out on the challenging data set called UCF-Crime. We select several abnormal behaviors in the data set for experiment, and for each abnormal behavior, we selects some videos in the dataset for experiment. After training, the model has greatly improved the detection effect of abnormal videos.