结合个性化建模和深度数据的三维人体姿态估计

利用深度传感器估计三维人体姿态是计算机视觉领域的一个重要问题，在人机交互、虚拟现实和动画设计等领域有重要的应用价值.针对该问题的主流方法是自底向上的方法，这类方法一般采用分类、回归或检索技术，可以直接从深度数据中估计三维肢体姿态，在人机交互中得到了很广泛的应用.但是这类方法依赖于大规模的姿态数据库，而且结果不够精确.本文提出一种结合个性化人体建模和深度数据的三维姿态估计方法，首先对运动对象建立三维虚拟人模型，然后利用该个性化的虚拟人模型与深度数据之间的点匹配关系构造姿态优化的目标函数，通过迭代优化目标函数，估计出与深度数据相吻合的三维姿态.与传统方法相比，本文方法不需要任何姿态数据库.实验表明，本文方法得到的结果更加精确.

Estimate 3D Human Poses from Personalized 3D Human Model and Depth Data

3D human pose estimation using depth sensor is an important research topic in computer vision, and it is useful for applications in human-computer interaction, virtual reality, and design of animation etc. The most successful methods toward this problem are bottom-up methods which predict 3d poses using classification, regression or retrieval techniques. These methods are widely applied in human-computer interactions. However, these methods rely on a huge human pose database and the predictions are rather inaccurate. In this paper, we propose to estimate 3D human pose using personalized 3D human models and monocular depth images. We firstly reconstruct a 3D virtual human model for each subject, and in the pose estimation phase, we reconstruct incomplete mesh from depth data, and estimate the correspondences between points of the 3d human model and the incomplete mesh. We estimate the optimal 3D poses through iterative optimization of objective function. In comparison with bottom-up methods, our method is free of any pre-captured dataset. Our experiments verifies that our results are more accurate than those of other methods.