Hand Pose Estimation from Depth Image Based on CPM
Jianming Yang, Pandi Li, Xiaowei Tu, Runchao Yu, Qinghua Yang · 2022 IEEE 10th Joint International Information Technology and Artificial Intelligence Conference (ITAIC) · 2022
Hand posture estimation using RGB images is often ineffective due to the relative position of hands and the background light. Depth picture contains the distance information of an object from the viewpoint, which can more accurately reflect the relative position of the object in the picture. In this paper, we first locate the hand position in the figure by centroid detection method, and then use Convolutional Pose Machines (CPMs) based on the residual network (ResNet) to estimate the key point position and perform pose mapping to complete the 3D reconstruction of the hand pose. Finally, a low-cost Kinectv2 camera is used for real-time detection, and the results demonstrate that the model has good real-time performance and accuracy for hand pose estimation in depth images. The average endpoint error (EPE) is 7.37mm and 10.3mm for the TCVL and MSRA datasets, respectively. In the 3D PCK test, the probability of correctly estimating the key point can reach 98% and 95% or more when the threshold is around 30mm (different thresholds for different test sets), which is superior to other mainstream algorithms.