Abstract: A novel method for reconstructing 3D motion of human avatar from real-time orthogonal images by using artificial neural network techniques was proposed. The input vector to the network was constructed by using the extracted coordinates of the feature points, while the output one indicated the 3D coordinates of the representative points and joints of the real human. The fitting process of the neural network was based on some proper neural network learning techniques with a set of sample data pairs that were obtained by using a motion capture system ReActor. The proposed method was implemented on a personal computer and ran in real-time applications. And experimental results confirm both the feasibility and the effectiveness of the proposed method for estimating 3D human motion (reconstruction error in MSE is less than 5%).

Key words: orthogonal images, 3D motion reconstruction, neural network

摘要： 提出一种由目标的立体图像通过人工神经网络实时估计得到其3D姿态的方法。网络的输入向量由同步立体图像帧上目标特征点的坐标构成；而输出向量则表示目标若干关键位置的三维姿态（进而可以建立目标的3D模型）。拟合该神经网络所需要的输出样本数据由运动捕获系统REACTOR获取。实验表明基于该算法的3D姿态估计误差低于5%，可以有效应用于3D虚拟目标的计算机实时合成等。

关键词: 立体图像, 三维重建, 人工神经网络