Learning deep event models for crowd anomaly detection

Abnormal event detection in video surveillance is extremely important, especially for crowded scenes. In recent years, many algorithms have been proposed based on hand-crafted features. However, it still remains challenging to decide which kind of feature is suitable for a specific situation. In addition, it is hard and time-consuming to design an effective descriptor. In this paper, video events are automatically represented and modeled in unsupervised fashions. Specifically, appearance and motion features are simultaneously extracted using a PCANet from 3D gradients. In order to model event patterns, a deep Gaussian mixture model (GMM) is constructed with observed normal events. The deep GMM is a scalable deep generative model which stacks multiple GMM-layers on top of each other. As a result, the proposed method acquires competitive performance with relatively few parameters. In the testing phase, the likelihood is calculated to judge whether a video event is abnormal or not. In this paper, the proposed method is verified on two publicly available datasets and compared with state-of-the-art algorithms. Experimental results show that the deep model is effective for abnormal event detection in video surveillance.