Anomaly detection is an important topic for surveillance video analysis and public security management. One of the major challenges comes from the fact that there is no abnormal data for training in most cases. Gaussian modelling has proven to be one of the most successful approaches to solve this one-class classification problem. Existing algorithms load features of all the training data and learn the Gaussian model in an offline way, which consumes a lot of memory and training time. Besides, they cannot handle the normal streaming data with varying patterns over time in real scenarios. In this paper, we propose an anomaly detection algorithm with passive aggressive online Gaussian model estimation. The algorithm is able to reduce the memory occupation and training time significantly without loss of model discriminability. The online learning strategy can also well adapt to the varying patterns. According to the experiments, the proposed algorithm can cut off over 99% memory occupation and 80% training time consumption.
展开▼
