科研成果详情

In the paper, we develop an unsupervised learning approach for anomaly event detection in videos based on a 3D ConvNet encoder-decoder for extracting spatial features and a ConvLSTM encoder-decoder for learning the temporal evolution of the spatial features. Moreover, squeeze-and-excitation networks (SENet) is incorporated into our model to take global information of each frame into account. In training, our model only includes normal events of video, whereas in testing, the videos have both normal events and abnormal events. The effectiveness of the proposed approach for anomaly event detection is validated through experiments on the UCSD datasets.