中文摘要：

对机器人跟踪手势姿态运动准确性的估计，能够有效提高人机交互效率。对手势姿态运动准确性的估计，需要通过非线性手势模型中间层，从而选取关节点处一个旋转自由度，完成对手势姿态运动准确性的估计。传统方法先对手势特征的提取，计算手部轮廓曲率，但忽略了关节点处旋转自由度的选取，导致手势姿态运动准确性估计精度偏低。提出将手势姿态运动准确性估计应用于手势识别中。通过应用卷积神经网络对深度图中人手进行姿态估计，得到关键关节点的空间坐标，计算出各手指弯曲程度进行猜拳手势识别，使得机器人跟踪手势姿态运动准确性图像可以同典型手势一样被识别。其中用于姿态估计的卷积网络模型，通过加入非线性手势模型中间层，选取关节点处一个旋转自由度，得到附加的旋转变换矩阵。完成对机器人跟踪手势姿态图像运动准确性估计。实验证明，上述网络模型使手势姿态估计中平均误差降低了2．21mm。

英文摘要：

A motion accurate estimation of gesture posture for robot tracking is proposed. Firstly, convolutional neural network is used to carry out posture estimation for human hand in depth map, and space coordinate of key artic- ulation point is obtained. Then, degree of crook of each finger is worked out to carry out gesture recognition of finger -guessing game. Thus,like typical gesture, the motion accuracy of robot tracking gesture posture image can be rec- ognized. A rotational degree of freedom at articulation point is selected via adding interlayer of nonlinearity gesture model, and additional rotation transformation matrix is obtained. Finally, the motion accurate estimation is completed. Experiment results show that the proposed method can reduce average error of gesture posture estimation by 2. 21mm.