Anomaly detection based on superpixels in videos

Based on superpixels, we propose a novel method for detecting abnormal events in videos. The conventional methods divide the frames into regular grids and consider the grids with low probability as abnormal events. By contrast with traditional approaches, we divide frames into superpixels according to their similarity and compactness, and the superpixels under the scene model mask are used as the anomaly candidates. The anomaly detection is carried out at two scales: the basic grids covered by the superpixel candidates and the actual superpixel itself. Anomaly scores are calculated by comparing the test samples with corresponding templates. Experiments on the public databases show that our method can effectively detect abnormal events in complex scenes.