Reiss Tal ; Hoshen Yedid

Abstract

Video anomaly detection (VAD) identifies suspicious events in videos, whichis critical for crime prevention and homeland security. In this paper, wepropose a simple but highly effective VAD method that relies on attribute-basedrepresentations. The base version of our method represents every object by itsvelocity and pose, and computes anomaly scores by density estimation.Surprisingly, this simple representation is sufficient to achievestate-of-the-art performance in ShanghaiTech, the most commonly used VADdataset. Combining our attribute-based representations with an off-the-shelf,pretrained deep representation yields state-of-the-art performance with a$99.1\%, 93.7\%$, and $85.9\%$ AUROC on Ped2, Avenue, and ShanghaiTech,respectively.