View Stabilized Skeleton Joint Descriptor for Human Action Recognition from Skeleton Joints

Human Action Recognition (HAR) through 3D Skeleton Joints has obtained substantial applicability in Human Computer Interaction and visual surveillance. But, the view point variations create a big challenge in designing a view invariant HAR system. Towards such intention, this paper proposes a new Action descriptor called as View Stabilized Skeleton Joint Descriptor (VS
²
JD) which transforms the skeleton joints from Cartesian coordinate system into a new view invariant coordinate system. VS
²
JD considers torso joints as reference joints and transform all the skeleton joints with respect to three (
$x, y$
and z) orientations. Further, the redundant joints are discarded after computing their weights through accumulated motion energy in an action sequence. Simulations are carried out on the large sized NTURGB+D dataset which consist of totally 60 different actions. On an Average, the proposed system gained an accuracy of 91.3020% and 86.3360% under cross view and cross subject evaluation protocols. The comparison between proposed and recent existing methods proves the superiority.