3-D Shape and Pose Estimation of Articulated Objects From Image Sequences

为了改善传统的手势界面工具在人机交互中的不便，本文研究了一种新的非触觉手势识别方法，该方法从摄像机拍摄的图像序列中估计人体的三维形状和运动，而不需要在用户的身体上做任何特殊的佩戴或标记。首先，研究了一种从单目轮廓图像序列中估计手形和姿势的方法。该方法首先给出一个粗略的形状模型，然后计算出手指长度等模型参数的概率分布，然后计算出模型参数的概率分布。该方法首先给出一个粗略的形状模型，然后计算出模型参数的概率分布，如手指长度，然后计算出模型参数的概率分布。该方法首先给出一个粗略的形状模型，然后计算出手指长度等模型参数的概率分布，然后计算出模型参数的概率分布，手掌大小和关节角度。结合每个图像帧中观察到的图像特征和关于手部的先验知识，如关节角度的可能范围和每个手指长度之间的关系，更新分布，更精确地修正模型形状。该估计算法在由三维高速图形生成器和个人计算机构建的平台上实现，并对计算机图形图像和真实手图像进行了实验。结果表明，如果获得不同形状的手的多个视图，则可以从单目图像中很好地估计关节角度、手指长度和手掌大小。由于轮廓包含的关于三维形状的信息非常有限，研究了一种利用图像运动和深度信息的方法。通过对该方法的精确验证，当图像特征的精度不高时，由于概率计算中的近似误差而导致形状和姿态的错误估计。发现需要其他不使用概率的方法来解决这一问题。其次，测试了另一种将可能的形状参数区域近似为椭球体的估计方法，以替代概率方法。结果表明，即使图像特征是错误的，该方法也可以正确地估计形状和姿态。然而，也发现该方法在突然的大误差输入下非常弱。

In order to improve inconvenience of a conservative gesture interface instrument for human-machine interaction such as "wired-glove", we studied the newly non-tactile gesture recognizing method which estimates the 3-Dimensional shape and movement of human body from image sequences captured by TV-cameras without any special wears or marking on a user's body.First, a hand shape and pose estimation method from monocular silhouette image sequences was studied.Given a roughly shaped model initially, the method estimates rough hand poses.Then it calculates a probability distribution of the model parameters such as finger lengths, palm size and joint angles.Combining observed image features in every image frame and prior knowledge about human hand such as possible ranges of joint angles and relationship between each finger lengths, the distribution is updated and the model shape is modified more precisely.This estimation algorithm was implemented on a platform constructed from a 3-D highspeed graphics generator and personal computers and experimented for comuter graphics images and real hand images.It was demonstrated the joints angles, finger lengths and palm size can be well-estimated from a monocular images if many views of various shaped hand are obtained.Since silhouettes contain very limited information about 3-D shape, a method using image motion and depth information was also studied.By precise verification of this method, the shape and pose are wrongly estimated due to approximation errors in probability calculation if accuracy of the image features isnot good.It found other method not using probability is required to resolve the problem.Next another estimation method which aproximate the possible shape parameter region as a ellipsoid was tested for subsutitution of the probablistic method.It was found this method can correctly estimate even if the image features are errornous.However, it was also found this method is so week in a sudden large errornous input.