Action pose recognition from 3D camera data using inter-frame and inter-joint dependencies

We developed a classifier for back stretching bodyposes used in therapy. Joint coordinates were collected and normalized to adjust for different body proportions and speeds. Initial classification on the entire normalized data set established a baseline of 25% accuracy, relative to 17% for chance. A second stage classifier was developed to use inter-frame movement kinematics to isolate the core frames where the final static pose was achieved. Combined with higher-order features from interjoint dependencies, the modifications produced an accuracy of 63%.