Learning Spatiotemporal Representation Based on 3D Autoencoder for Anomaly Detection.

Because of ambiguous definition of anomaly and the complexity of real data, anomaly detection in videos is of utmost importance in intelligent video surveillance. We approach this problem by learning a novel 3D convolution autoencoder architecture to capture informative spatiotemporal representation, and an 2D convolutional autoencoder to learn the pixel-wise correspondences of appearance and motion information to boost the performance. Experiments on some publicly available datasets demonstrate the effectiveness and competitive performance of our method on anomaly detection in videos.