View-invariant Feature using Pose Information and Flexible Matching Algorithm for Action Retrieval

Action retrieval and detection utilizing pose in-formation has two difficulties in industrial applications, the occlusion and variation of human actions. To overcome these difficulties, we propose a new normalization method to generate view-invariant pose feature that is corresponding to the key-points of the original pose information, and design flexible matching algorithms to perform high-accurate action retrieval. The proposed methods enable users to easily and flexibly perform action retrieval with weight constraints on specific body parts or neglect of invisible body parts for similarity computation. The experimental results are reported to show that our feature and flexible matching algorithm outperforms the state-of-the-art methods in a simulated dataset with annotated multi-view 2D poses and a real-world video dataset.