【作者】 潘华伟；

【导师】 孙家广； 李仁发；

【作者基本信息】 湖南大学 ， 计算机应用技术， 2013， 博士

【摘要】 人体运动姿态捕获与姿态分析是计算机视觉领域研究的重点问题之一,在影视游戏、监控分析、人机交互、虚拟现实、医疗诊断和运动分析等方面具有重要的应用价值,具有广阔的应用前景。当前,由于存在非刚性的人体形状、二维到三维投影的歧义性、自遮挡、高维参数的恢复、真实场景下图像特征提取与匹配等诸多困难,从数字视频图像中恢复出人体三维运动姿态存在大量的理论问题。因此,研究人体姿态捕获与姿态分析既具有理论研究意义,又具有工程应用价值。人体姿态重建与跟踪是人体姿态分析的基础之一。本文围绕人体姿态重建及跟踪开展研究,包括人体与手势的重建与跟踪,研究了在多相机环绕环境下的人体定位和姿态跟踪算法和方法,可以为面向自然真实场景的人体姿态捕获与姿态分析提供理论基础和系统框架。具体地,本文的主要研究内容和创新点如下：第一、针对复杂场景下的前景目标提取,提出了一种结合深度信息和肤色信息的分割方法。在具有复杂运动背景的环境中提取前景目标时,利用深度信息可以帮助避免复杂背景和背景物体运动的干扰。本文把RGB相机和Depth相机二者相机结合,提出一种把深度信息和肤色信息结合的分割方法,深度信息可以帮助避免复杂背景的干扰。在背景复杂或有干扰的情况下,提取的前景效果相对于传统方法来说更为健壮有效。第二、针对姿态构造的特点,结合层次策略,提出了一种间接人体模型的自动初始化方法。传统方法通常使用学习方法或已知参数的固定姿态,以及形状控制参数、层次策略等方法构造肢体骨架。本文结合层次构造法,提出了基于矢量合成分析方法和改进细化体素方法,构建了具有与真人匹配的人体骨架和拓扑结构,以自适应的模型来替代普遍模型。第三、针对一个旋转角分解为分别绕XYZ轴的旋转角时,传统方法是必须假设其中一个旋转角度为零,本文提出了一种引入构造分解法求解旋转自由度的方法。一个空间点相对于某点的旋转角,如果要分解为相对于XYZ三轴的旋转角度,通常需要计算出绕XYZ三轴的旋转分量,传统方法必须假设其中一个旋转角度为零,通过降维来求解另外两个旋转分量,该过程不能真实描述各关节点的运动情况。为了解决此问题,本文引入构造分解法,R是正交矩阵,存在U和Q,使得正交阵R=QUT,选取适当矩阵元素,可使正交矩阵R分解为分别绕Z轴、X轴和Y轴的旋转矩阵的乘积,从而有效地表达了各关节点的旋转信息。第四、针对传统运动模型自由度较高的问题,提出了新的运动模型,降低了自由度维数,并利用此模型实现了运动姿态跟踪。合理利用人体生理运动原则及非标记运动的特点,对运动模型重新设计,将运动自由度维数降至最低,使得自由度的求解速度加快。提出使用一种运动囊概念,运动囊反向驱动骨架的节点,节点必须在位移和角度上跟随变化,通过使用多线索跟踪技术,把运动囊轮廓边界线索和重合度线索结合,用梯度法计算姿态自由度近似最优值,进而更新人体骨架各节点的姿态参数。第五、基于前面提出的算法,开发出了一套新的基于姿态捕获的三维重建原型系统。设计了一维标定物对相机内外参标定,标定方法快捷实用。对前面各章方案的有效性进行分析和验证。系统简单实用、成本较低。更多还原

【Abstract】 The capture of human body posture and its posture analysis is one of the hottest research topics in the field of computer vision. It has great potential in various applications such as movies&games, surveillance analysis, human machine interaction, virtual reality, medical diagnosis and motion analysis. However, there are still many difficulties such as the non-rigid body shape, the ambiguity of2D to3D projection, self-occlusion, recovery of high-dimensional parameters and the feature extraction and matching under real situations. There are still many theoretic issues about the recovery of human body posture from digital video. Therefore, the research on human body posture analysis is of both theoretic and practical value.Reconstruction and analysis of human body posture is one of the bases of human body posture analysis. This thesis focuses on human body posture analysis, including the reconstruction and tracking of human body and hand posture. It researches on the human body location and posture tracking under multi-camera environment, which provides theoretic basis and systematic framework for human body posture capturing and analysis. In detail, the main work and contributions are summarized as follows.First, a novel algorithm which combines depth and skin information and can extract fore-ground target object in the complex scene is proposed. To segment the fore-ground image in the complex scene, the depth segmentation method is an effective method to avoid the interference in the complex scene and moving objects in the background. With the RGB camera joint with Depth camera, a segmentation method combined with depth information and skin color information is proposed. The depth information help to get clear fore-ground target object. The method is robust and efficient even in the background there have the complex scene and moving object.Second, according to the characters of pose structure, and combining the hierarchy strategy, an auto-initialized algorithm which used in indirect model human model is proposed. The general methods retrieving the pose parameters are matching a fixed pose prepared beforehand which has known the parameter, or using the learning method, or shape controlling parameters, or hierarchical strategy. We combine hierarchical strategy to propose Vector Synthesis Analysis (VSA) method and improve the voxel thinning method. The auto-initialized algorithm builds the human skeleton and topology which is matching the real man. We construct adaptive model method instead of general method.Third, when a rotation angle is decomposition three sub-angles around XYZ axes respectively, the traditional methods is to make a hypothesis that one of the sub-angles must be zero. A construction of decomposition method which is used in solving the problem is proposed in the paper. If a rotation angle that a point rotates around another point need to be decomposed into three sub-angles around axes XYZ respectively, the general method is to reduce the number of dimensions that make a hypothesis that one of the sub-angles is zero, and then solve the other two rotation sub-angles. The process is not exactly describing the motion of the joints. In order to solve the problem, we introduce decomposition method. The rotation matrix R is orthogonal matrix. There are existing matrixes U and Q to make R=QUT Selecting the appropriate matrix elements, the decomposition method can decompose arbitrary rotation angle into combination of three sub-angles rotating around XYZ axes. The method effectively represents the rotation information of each joint.Fourth, according to traditional motion model that exist the problem of high dimension of Degree of Freedom (DOF), a new motion model which has the lower dimension of DOF is proposed, and we use it to implement the posture parameters recovery. Rationally using the physiological motion principles and the characters of markless motion, we redesign the human motion model and decrease the quantity of DOF to lower, which make help accelerating to recover the pose parameters. And we also use a motion capsule concept. The motion capsule inversely drives the joints of skeleton which have to change with translation and rotation angle. We use multi-cue tracking technology which combined with silhouette cue and overlap cue of motion capsule. We used local gradient-based method to recover the optimal parameters and update the pose with them.Fifth, a framework of markless human posture capture system is constructed based on above-mentioned work. We designed a1D calibration rod to calibrate the intrinsic and extrinsic parameters which make calibration easier and quicker. And we construct a simple, practical, low-cost prototype system at last.更多还原