Anomaly detection with multi-scale pyramid grid templates.

In this paper, we propose a method for abnormal event detection in videos based on Multi-scale Pyramid Grid Templates (MPGT). Unlike traditional methods that usually finish anomaly detection based on a single scale feature, we propose to detect anomalies with a designed multi-scale normalized motion feature in the framework of MPGT. In our work, two scene models are proposed, including a global model and an online model, because the anomalies often occur in regions with moving objects. In addition, we propose a fast method for computing high-scale motion features using a convolution operation based on the first scale feature, and design a scheme for combining the detection results at different scales using vote and pyramid strategies. Experiments on public datasets show that our method has a balanced performance on all the testing datasets. [ABSTRACT FROM AUTHOR]