Study of an Action Recognition Method Using a Human Partial Skeleton Model Extracted from Three-dimensional Time-Sequential Images

Abstract

This paper proposes a method that recognizes the grasping action of the upper-half of the human body from RGBD time-sequential images in case that an object is placed at different positions on a desk and that the object is grasped by the human in different manners. The proposed method consists of modules for extracting human body features and for recognizing human actions. In the feature extraction module, to Vemulapalli et al.’s twenty features based on relative poses between the links connecting human joints, five features concerning the neck as well as the thumbs and finger tips of both hands are newly added. In addition, a partial skeleton model that uses only the features in the upper-half of the body is utilized. The action recognition module uses a linear-SVM (Support Vector Machine) for each action to be recognized. Experiments for recognizing 10 actions that include nine combinations of three grasping manners and three different object positions as well as “do nothing” show promising recognition accuracies of the proposed method.