Learning Spatiotemporal Features With 3DCNN and ConvGRU for Video Anomaly Detection

Video anomaly detection aims to analyze the abnormal events or behaviors from massive monitoring video data, which is extremely challenging due to the ambiguous definition of abnormal behavior and the complex monitoring scene. Feature representation based on the hand-crafted of video local spatial area is more complicated, and it is difficult to learn the essential feature from the input video. In this paper, a deep autoencoder network combined with 3DCNN and ConvGRU is proposed to learn the spatiotemporal features for video anomaly. Firstly, 3DCNN and bidirectional ConvGRU are used to encode the local-global spatial features and short-long-term temporal features in the spatiotemporal dimension. Secondly, the reconstruction branch is introduced to reconstruct video frames, while the prediction branch is utilized to make the encoder to learn the better spatiotemporal feature at the training phase. In addition, the regularization of adjacent frames in a loss function is carried on to improve the temporal feature. The weights of the C3D model trained by action recognition are transferred to 3DCNN to prevent model overfitting. Experiments on real anomaly datasets shows the effectiveness of our proposed deep model.