Normalizing Flows for Human Pose Anomaly Detection

Video anomaly detection is an ill-posed problem because it relies on manyparameters such as appearance, pose, camera angle, background, and more. Wedistill the problem to anomaly detection of human pose, thus decreasing therisk of nuisance parameters such as appearance affecting the result. Focusingon pose alone also has the side benefit of reducing bias against distinctminority groups. Our model works directly on human pose graph sequences and isexceptionally lightweight (~1K parameters), capable of running on any machineable to run the pose estimation with negligible additional resources. Weleverage the highly compact pose representation in a normalizing flowsframework, which we extend to tackle the unique characteristics ofspatio-temporal pose data and show its advantages in this use case. Thealgorithm is quite general and can handle training data of only normal examplesas well as a supervised setting that consists of labeled normal and abnormalexamples. We report state-of-the-art results on two anomaly detectionbenchmarks - the unsupervised ShanghaiTech dataset and the recent supervisedUBnormal dataset.