Object-Centric Video Anomaly Detection with Covariance Features.

In this paper, we propose four different methods for object-centric anomaly detection in surveillance videos based on autoregressive probability estimation. By means of the methods we propose, normal (typical) events in a scene are learned in a probabilistic framework by estimating the features of consecutive frames taken from the surveillance camera. To decide whether an observation sequence (i.e. a small video patch) contains an anomaly or not, its likelihood under the modeled typical observation distribution is thresholded. Due to its effectiveness in object detection and action recognition applications, covariance features are used in this study to compactly reduce the dimensionality of the shape and motion cues of spatiotemporal patches obtained from the video segments. By employing an object detection module to determine the important active regions in a scene with high detection rate, we propose new long-short term memory (LSTM), linear regression, and Gaussian mixture based methods to model the probability density of observation sequences. The most successful methods we propose achieves an average performance of 0.843 and 0.935 AUC scores respectively on two publicly available benchmark datasets.