Unsupervised Abnormal Crowd Activity Detection Using Semiparametric Scan Statistic

We propose a fully unsupervised method for abnormal activity detection in crowded scenes. Neither normal nor abnormal training examples are needed before detection. By observing that in crowded scenes, normal activities are the behaviors performed by the majority of people and abnormalities are behaviors that occur rarely and are different from most others, we propose to use a scan statistic method to solve the problem. It scans a video with windows of variable shape and size. The abnormality of each window is measured by a likelihood ratio test statistic, which compares two hypotheses about whether or not the characteristics of the observations inside and outside the window are different. A semiparametric density ratio method is used to model the observations, which is applicable to a wide variety of data. To reduce the search complexity of the sliding window based scanning, a fast two-round scanning algorithm is proposed. We successfully applied our algorithm to detect activities that are anomalous in different ways, achieving performance competitive to other state-of-the-art methods which requiring supervision.