Abnormal Event Detection and Localization in a Video based on Similarity Structure

This paper introduces a method for abnormal event detection in video. The video is divided into a set of cubic patches. A new descriptor for representing the video patches is proposed. This descriptor is created based on the structure similarity between a patch and nine neighboring patches of it. All training normal patches in respect to the proposed descriptor are represented and then modeled using a Gaussian distribution as the reference model. In test phase, those patches which are not fitted to the reference model are labeled as anomaly. We have evaluated the proposed method on two UCSD[1] and UMN[2] popular standard benchmarks. The performance of the presented method is similar to state-of-the-art methods and also is very fast.