基于双循环Transformer的三维人体姿态估计

3D Human Pose Estimation Based on Dual Circulation Transformer

针对视觉传感器采集到的图像进行三维人体姿态估计,提出一种双循环Transformer网络模型,有效地从二维关键关节点中提取时空维度高相关性特征,增大感受野,从而提高三维姿态估计的精度。通过在视觉传感器采集得到的公开数据集Human3.6M上的仿真实验,验证了双循环Transformer算法的性能。分析结果表明,最终估计得到的三维人体关节点的平均关节点位置偏差MPJPE(Mean Per Joint Position Error)为41.6 mm,相比于现有方法有一定提升,可以应用到许多下游相关工作中,有着较强的应用价值。

A dual-circulation Transformer network is proposed for 3D human pose estimation from video captured by visual sensors,which is capable of effectively extracting highly correlated features in the spatio-temporal dimension from 2D key joints,thereby increasing the perceptual field and improving the accuracy of 3D pose estimation.The performance of the dual-circulation Transformer algorithm is verified through simulation experiments on the public dataset of Human3.6M,obtained from visual sensor captures.The analysis results reveal that the dual-circulation Transformer algorithm achieves a mean per joint point position error(MPJPE)of 41.6 mm for the final estimated 3D human joint points,demonstrating competitive detection results.Moreover,the proposed approach exhibits strong potential for application in various downstream related works.