A novel statistical learning-based framework for automatic anomaly detection and localization in crowds

We propose a novel framework for fast and robust video anomaly detection and localization in complicated crowd scenes. Images of each video are split into cells for extracting local motion features represented by optical flow. In the train videos, most background cells are subtracted by ViBe model. Feature vectors are extracted from each cell by integrating the value of optical flow in 8 different direction intervals. Then we apply Principal Component Analysis (PCA) to transform the feature vectors. The normal activity patterns in the train videos are learnt by constructing a Gaussian Mixture Model (GMM) upon the feature vectors. For any new feature vector extracted from the test video clips, we use the learnt model to calculate a probability value to represent normal level of each cell. Considering the continuity of the motion, we also use abnormal information obtained from previous frames as a supplementary for anomaly prediction in the current frame. At last, we determine whether an activity pattern of a cell is normal or abnormal by using mean shift to cluster the probability values of the frame. Qualitative experiments on real-life surveillance videos, the recently published UCSD anomaly detection datasets, validate the effectiveness of the proposed approach.