Recovery of body postures from a monocular image wequence based on probabilistic model learning

在视觉监控、福利机器人和人机界面领域，计算机能够识别人的姿势/运动是很重要的。在本研究中，我们尝试从单眼图像序列中估计人体姿势。我们提出了循环运动模型，这是一种表示运动的状态空间模型。在该模型中，假设运动是循环的，系统的状态由运动的相位表示。状态（即相位）是用顺序蒙特卡罗方法从单眼图像序列中估计出来的。然后，根据状态计算姿态。循环运动模型从运动捕捉数据中学习。此外，我们扩展了循环运动模型，以允许相机视点的变化。实验证明了该方法的有效性。

In the fields of visual surveillance, welfare robotics, and human interface, it is important that a computer can recognize the posture/motion of people. In this research, we try to estimate human postures from monocular image sequences.We propose the cyclic motion model which is a state space model to represent motions. In the model, the state of a system is represented by the phase of a motion, assuming that the motion is cyclic. The state (i.e. the phase) is estimated from a monocular image sequence using the sequential Monte Carlo method. Then, the posture is computed from the state. The cyclic motion model is learned from motion capture data. In addition, we extend the cyclic motion model to allow for camera viewpoint changes.Experiments show the validity of the proposed technique.