Detecting Anomalies in Images by using an Autoregressive Model

摘要

We propose an. anomaly detection algorithm for images. The proposed algorithm is based on an autoregressive model. In practice, we employ PixelCNNs to estimate the likelihood of images according to training sets of data. The detection performance of the proposed system is evaluated on a homemade dataset of surveillance videos recorded by a drone, as well as several public benchmarks of surveillance videos recorded by fixed-position.