Anomaly event detection and localization of video clips using global and local outliers

The automatic detection of anomaly events in video sequence has become a critical issue and essential demand for the extensive deployment of computer vision systems such as video surveillance applications. An anomaly event in video can be denoted as outlier behavior within video frames which formulated by a deviation from the stable scene. In this paper, an anomaly event detection and localization method in video sequence is presented including multilevel strategy as temporal frames differences estimation, modelling of normal and abnormal behavior using regression model and finally density–based clustering to detect the outliers (abnormal event) at clips level. Hence, outlier score is obtained at the segment or clip level along video frames sequences. The proposed method seplits video frames into nonoverlapped clips using global outlier detection process. Afterward, at each clip, the local outliers are determined based on density of each clip. Extensive experiments were conducted upon two public video datasets which include dense and scattered outliers along video sequence. The experiments were performed on two common public datasets (Avenue) and University of California, San Diego (UCSD). The experimental results exhibited that the proposed method detects well outlier frames at clip level with lower computational complexity comparing to the state-of-the-art methods.