Posture estimation from Kinect image using RVM regression analysis
Hiroyuki Fujimura, Hyoungseop Kim, Joo Kooi Tan, Seiji Ishikawa · 2015
Kinect is always used as a device to estimate posture. However, there are difficult to estimate the posture in the case of using a Kinect only. Therefore, we propose a method to estimate more accurately posture by synthesizing the posture obtained by Kinect and estimated by the regression analysis. In the regression analysis, we associate the HOG features and joint parameters that consists of 20 coordinate points. Posture data used for learning of the regression model is used difficult posture be obtained with Kinect. Similarity in brightness between frames at around each joint of the skeleton obtained by regression analysis and Kinect is calculated. Then we synthesize the posture by calculating a weighted average. In addition, RVM regression model is used to improve the accuracy of representing the posture by the proposed method.