Graph Embedded Pose Clustering for Anomaly Detection, ArXiv'19 {paper} {code}

Amir Markovitz, Gilad Sharir, Itamar Friedman, Lihi Zelnik-Manor, and Shai Avidan

Objective

Given examples of only normal actions, be able to classify normal and abnormal actions

Learn a 'bag-of-words' representation of actions using clustering.

Estimate normality given distances to clusters.

Method

Assume working pose estimation from video.

Temporal pose as input (keypoint positions over time)

Soft clustering of actions

Experiments

• ShanghaiTech Campus dataset anomaly detection benchmark

  • Normal behaviour: walking, abnormal is running, biking, ...

• various dance styles are 'normal', other motion is abnormal

  • On NTU-RGB+D where 3D keypoints are extracted using kinect
  • 250 out of 400 actions in Kinetics400
    • subset of action as 'normal', rest 'abnormal'

'Fine-grained' --> only one action is normal

'Coarse-grained' --> many action classes are normal