Unsupervised Abnormality Detection in Video Surveillance

Abstract

{nanri, otsu} @ isi.imi.i.u-tokyo.ac.jp The detection of abnormal movements is an important problem in video surveillance applications. We propose an unsupervised method for abnormal movement detection in scenes containing multiple persons. Our method uses cubic higher-order local auto-correlation (CHLAC) to extract movement features. We show that the additive property of CHLAC in combination with a linear subspace method is well suited to simplify the learning of normal movements and to detect abnormal movements even in scenes containing multiple persons. One particular advantage of this method is that it does not necessitate the object segmentation and tracking and also any prior knowledge about objects. Some experimental results are shown to exhibit the validity of the method. 1