融合深度学习与外极线约束的三维人体位姿测量方法

为了实现汽车座椅上三维人体姿态的高精度实时测量，基于现有二维人体关节点和三维人体关节点测量方法，提出了一种融合深度学习与外极线约束的三维人体姿态测量方法。该方法将二维人体关节点深度网络提取方法和双目测量系统相结合，采用双通道多阶段迭代网络分别提取左右相机图像中人体二维关节点，结合关节点位置的Brief特征和外极线约束，利用双目相机标定结果将匹配二维关节点信息转换到三维空间中，最终得到三维人体姿态。实验结果表明，文中提出方法在自采测试集中的检测精度可达到98%。通过得到三维关节点计算所得关键位姿角度的偏差小于10°。该文所提出的方法能够满足实际汽车座椅设计的数据采集要求。

3-D human pose measurement method based on deep learning and epipolar constraint

It is necessary for the design of car seat to realize the high precision and real-time measurement of the 3-D human pose on the car seat. Based on the existing 2-D and 3-D human joint measurement methods, a 3-D human pose measurement method is proposed, which integrates deep learning and epipolar constraint. This method combines 2-D human joint depth network extraction method with binocular measurement system, uses two-channel multi-stage iterative network to extract 2-D human joint points in the images taken by left and right cameras respectively, combines the Brief feature of joint point position and epipolar constraint, uses binocular camera calibration results to convert the matching 2-D joint point information into 3-D space, and finally obtains 3-D human pose. The experimental results show that the detection accuracy of the proposed method in the self sampling test set can reach 98%. The deviation of the angle of the key position of human body calculated by the 3-D joint point is less than 10°. Therefore, the method proposed in this paper can meet the data acquisition requirements of the actual car seat design.