DNNを用いた2Dイメージベース動作解析システムによる人体関節の3Dポジション検出

抄録

Motion analysis system has been used for making movies or TV games, improving sports skills and medical fields. The main types of motion analysis system are two methods with cameras and force plates or inertial sensors. Former is generally called “an optics type” and latter is “a wear type”. Considering analyzing a ski player on the snow ground, we cannot apply those motion analysis methods. The purpose of this study is to detect 3D human joints from 2D ski player in a movie. Therefore, we prepared the “3D solid model” for a motion analysis method. Joints of 3D model can be freely moved on 3DCAD system. And we proposed the solution that we can get three-dimensional motion by superimposing 3D solid model on a 2D ski player in a picture. This work is called “matching process” in our study. We can analyze ski player motions with proposed method. However, it was 19.6 hours in one turn to superimpose them because we have to apply“matching process” in an every flame of motion movie. Therefore, we prepared some templates have some poses to shorten the matching work time and we selected the appropriate template using Deep Neural Network (DNN). We reviewed the process, to we can automate for speed up after automatic template selection with DNN.