This paper proposes a real-time abnormal behavior detection using Conditional Random Fields(CRFs). A normal behavior can be characterized by the spatial and temporal features obtained from the video of human activities. The difficult of abnormal behavior detection is that human behavior varies in both motion and appearance. It is a continuous action stream, interspersed with transitional activities between abnormal and normal events. Here, we propose Bag of Words (BoWs) to describe the motion information as the observations. Then, we apply the CRFs and adaptive thresholding to identify the abnormal behaviors. Different from previous methods, our method can identify the undefined and unknown abnormal activities.
