Incident Detection Based on Dynamic Background Modeling and Statistical Learning Using Spatio-Temporal Features

This paper presents a method for detecting an incident motion (e.g., tumble or violent action of a person) in a dynamic background scene. This method is based on the use of spatio-temporal features obtained using a space-time patch (ST-patch). Our approach consists of three steps: 1) dynamic background modeling using a Gaussian mixture model, 2) human regions detection based on Real AdaBoost, and 3) calculation of irregularity measure using weighted ST-patch features. The proposed method can be used to detect the incident motion in a scene with a dynamic background, which would be difficult to detect with a conventional method using cubic higher-order local auto-correlation (CHLAC) features. Our experimental results show that our method performs about 27% better than the conventional method in a contained scene with an escalator as the dynamic background.