Fully Convolutional Network for 3D Human Pose Estimation from a Single View

• Main Authors: LIN, GUAN-HAN, 林冠翰
• Other Authors: LIE, WEN-NUNG
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/9gqs4c

碩士 === 國立中正大學 === 電機工程研究所 === 106 === This thesis called ”Fully Convolutional Network for 3D Human Pose Estimation from a Single View”. First we use RGB image from a single view as input and extract the 2D human pose from a deep-learning pre-trained model. Second, using deep-learning technique to estimate 3D anchor pose. Finally combined 2D human pose and 3D anchor pose as input into second deep-learning model to estimate 3D human pose. The 3D human pose estimation algorithm contained three steps. First using all of the 3D Human samples into K-means algorithm, and take the mean pose in each cluster as anchor pose ground truth in training part. Second, we trained first model called “3D anchor pose estimator”, which use 2D human pose as input ground truth, and the result of clustering as output ground truth. Anchor poses are some common human posture in our daily life. Final we train second model called “3D human pose estimator” and combine 2D human pose and 3D anchor pose to estimate final 3D human pose. In on-line test, we let images which captured from webcam as input into 2D human pose estimator to extract 2D human pose, and then path through our model to estimate 3D human poses. In this thesis, we will compare that using different normalization method with 2D pose samples and 3D pose samples, and analysis the influence with result. After adjust the parameters of the model, The error of each joint reduce 52% average. After samples normalization, our model can adapt more situation when camera moved or image resolution has been changed.