Video Anomaly Detection and Localization using Motion-field Shape Description and Homogeneity Testing

•We introduce a histogram-based shape descriptor to motion field in each local patch.•The motion descriptor captures the motion trend and details in local patches.•We propose a similarity-based statistical model to detect spatio-temporal anomalies.•The statistical model relies on unsupervised learning without any prior assumption.•The method can adapt to the whole scene with tolerance to perspective distortion.