The Algorithm Researching of Hand Pose Estimation Based on 3D Joint Regression
Ruimin Zhang, Shuqiang Du · 2022
The RGB image is used for 3D hand pose estimation, and the Gaussian function is defined according to the kinematics principle to represent the thermal map dependency of the previous joint point of the finger on the next joint point. The convolution network is assembled through the tree structure through the branching strategy, and the tree branch hierarchical network algorithm framework is designed. The two-dimensional pose of the five single finger joint points of the hand is estimated according to the local pose characteristics, and then the three-dimensional joint normalization regression algorithm is used to estimate the full three-dimensional hand pose for the two-dimensional pose of the single finger joint points. When regularizing linear regression, all features are punished. The degree of punishment is selected by the optimization standard of the sum of squares of residuals. The heat map loss function and joint point regression loss function are set. The experiment is carried out on the CMU Panoptic Dataset and STB Dataset. The experimental results show the advantages of this method by comparing parameters with other methods.