Abnormal Event Detection Based on Multi-scale Markov Random Field

In this paper, we present a novel unsupervised method for abnormal behavior detection, which considers both local and global contextual information. For the local contextual representation, we firstly divide video frames into local regions, then extract low-level feature such as histogram of orientated optical flow (HOF) and sequential feature which is composed of K temporal adjacent frames for each region. The global contextual feature encodes the statistical characteristics of those local features like orientation entropy and magnitude variance. An online clustering algorithm is introduced to generate dictionaries for the local and global features respectively. Then, for any new incoming feature, a maximum posterior estimation of the degree of normality is computed by multi-scale Markov Random Field (mMRF) based on the learned model. The proposed method is evaluated on hours of real world surveillance videos. Experimental results validate the effectiveness of the method, and the detection performance is promising.