基于双Kinect的人体运动捕捉

发布时间：2018-03-12 21:17

本文选题：运动捕捉　切入点：姿态估计　出处：《南京大学》2017年硕士论文　论文类型：学位论文

【摘要】：人体运动捕捉,就是从单个视角或者多个已同步或者非同步视角的数据或者序列中恢复出人体的姿态参数。精确性和鲁棒性是动态人体运动捕捉算法的基本追求。长期以来,很多学者和工程师都在致力于研究先进的人体运动捕捉算法,以期望能实时、自动、准确地构建三维空间中的人体运动形体。精确的人体运动捕捉,广泛应用于体育运动分析、远程医疗和虚拟现实等多个领域,丰富人们的娱乐方式和日常生活。当前,学术界在人体运动捕捉方面已经取得了巨大的进展,然而当遮挡发生时,当前比较流行的运动捕捉技术容易陷入局部最优,降低了运动估计结果的精确度,这促使着众多的学者和工程师更积极地投入到运动捕捉算法的研究中去。本文提出了一个完整的高精度的人体运动捕捉系统。首先,我们介绍了人体语义参数化的人体体型重建方法。本方法通过建立人体的三维模型与人体的语义参数(如身高、三围等)之间的线性映射关系,再利用采集得到的人体三维点云数据的约束,精确得重建人体体型模型。精确的人体体型模型使得人体运动姿态重建更加精确。对于运动捕捉,与以往的单目深度相机系统不同,我们使用两个Kinect相机从不同的角度来获取人体的运动信息,并以根据前一帧图像数据求解得出的人体形体作为当前待求解的图像的人体模板,将人体姿态参数的求解转化为一个MAP问题来解决。双Kinect相机系统的使用,使得当遮挡的情况出现时我们仍然有足够的信息用于精确估计人体的运动姿态。另一方面,利用人体运动在时域上的连续性,我们引入了时域约束条件,使用学习的方法来进一步挖掘人体姿态变化上的时域信息来约束人体运动参数的求解。时域约束条件的引入,使得我们在面对较快速度运动的人体时的求解结果精度更高。实验证明,时域约束条件有助于提高我们的运动捕捉系统的精度。
[Abstract]:Human motion capture is to recover human posture parameters from data or sequences of single or multiple synchronized or asynchronous visual angles. Accuracy and robustness are the basic pursuit of dynamic human motion capture algorithm for a long time. Many scholars and engineers are working on advanced human motion capture algorithms in the hope that they can construct human motion bodies in three dimensional space in real time, automatically and accurately. Accurate human motion capture is widely used in sports analysis. Telemedicine and virtual reality enrich people's entertainment and daily life. At present, great progress has been made in the field of human motion capture, but when occlusion occurs, The current popular motion capture techniques tend to fall into local optimum, which reduces the accuracy of motion estimation results. This has prompted many scholars and engineers to devote themselves more actively to the research of motion capture algorithms. In this paper, a complete human motion capture system with high accuracy is proposed. In this paper, we introduce the method of human body shape reconstruction with semantic parameterization. By establishing the linear mapping relationship between human body 3D model and human body semantic parameters (such as height, Sanwei, etc.), Then using the constraint of the collected 3D point cloud data, the body shape model can be reconstructed accurately. The precise body shape model makes the human body posture reconstruction more accurate. Different from the previous monocular depth camera system, we use two Kinect cameras to obtain human body motion information from different angles, and take the human body shape obtained from the previous frame image data as the human body template of the current image to be solved. The solution of attitude parameters of human body is transformed into a MAP problem. With the use of double Kinect camera system, we still have enough information to estimate the human body's motion attitude when the occlusion occurs. On the other hand, Using the continuity of human motion in time domain, we introduce the time domain constraint condition, and use the learning method to further excavate the time domain information of the human body attitude change to constrain the solution of the human body motion parameter, and introduce the time domain constraint condition. The experimental results show that the time-domain constraint condition is helpful to improve the accuracy of our motion capture system.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】