基于视频的三维人体重建和运动捕捉

发布时间：2018-03-25 16:37

本文选题：运动捕捉　切入点：摄像机标定　出处：《天津大学》2009年博士论文

【摘要】： 随着数字媒体技术的迅速发展,游戏、电影、动漫等产业对运动捕捉技术的需求增长很快,但是目前购买商业运动捕捉系统的成本非常高,使用环境要求严格,需要在演员身上安装标记点等特殊设备,非常不便。本文结合计算机视觉和图形学相关技术,提出一套新颖、实用、成本很低的三维人体模型重建和运动跟踪的方法。首先提出可以同时进行三维特征点重建和摄像机标定的迭代算法。该方法只需要少量的人工交互,算法收敛后可以得到精确的摄像机参数和重建模型上的三维特征点。其次,基于重建的三维特征点,使用径向基函数对通用模型进行变形,获得重建的中间模型。实验表明此方法可以适用于不同体型的对象,重建模型的质量较好,但是由于重建的特征点过于稀疏,人体的四肢对应不够好,所以模型还需要进一步改进。第三,为了提高重建模型的质量,提出以剪影匹配算法为基础的模型变形算法。该方法较准确的找到模型剪影和图像剪影之间的匹配点之后,使用径向基函数对中间模型进行变形,提高中间模型投影与图像的吻合程度,从而获得目标模型。实验结果表明目标模型在图像上的平均投影误差不超过1个像素。第四,为了解决因为图像解析度不够高和摄像机标定误差导致的局部走样问题,设计了三种过滤器:平滑过滤器、切片过滤器和法线过滤器,有效的解决了目标模型上的鼓包、扭曲变形等需要优化的问题。第五,基于图像的纹理映射通常因为光照差别较大出现纹理不连续问题。本文提出使用旋转切平面将模型切割,然后生成合成纹理的方法。该纹理映射到目标模型表面后,效果较好,克服了大部分纹理不连续的问题。最后,本文提出一种基于分析-合成的三维人体运动捕捉方法,实现了无标记点的基于复杂背景视频的人体运动捕捉。该方法对模型渲染图像和真实图像进行匹配,应用下山单纯形-模拟退火方法最小化匹配误差。实验结果表明此方法可以较准确稳定的进行运动捕捉。
[Abstract]:With the rapid development of digital media technology, the demand for motion capture technology in games, movies, animation and other industries has increased rapidly, but at present the cost of purchasing commercial motion capture systems is very high, and the use of environmental requirements is strict. It is very inconvenient to install special equipment such as marking points on actors. In this paper, a novel, practical and low cost method for 3D human model reconstruction and motion tracking is proposed, which combines computer vision and graphics technology. An iterative algorithm which can reconstruct 3D feature points and calibrate the camera simultaneously is proposed, which requires only a small amount of manual interaction. After the algorithm converges, the exact camera parameters and the 3D feature points on the reconstruction model can be obtained. Secondly, based on the 3D feature points of reconstruction, the general model is deformed by radial basis function (RBF), and the intermediate model is obtained. Experiments show that this method can be applied to objects of different shapes, and the quality of the reconstructed model is good. However, because the feature points of reconstruction are too sparse and the human limbs do not correspond well, the model needs further improvement. Thirdly, in order to improve the quality of the reconstructed model, a model deformation algorithm based on the silhouette matching algorithm is proposed, which finds the matching points between the model silhouette and the image silhouette more accurately. The radial basis function is used to deform the intermediate model to improve the degree of consistency between the projection of the intermediate model and the image. The experimental results show that the average projection error of the target model on the image is less than 1 pixel. Fourthly, in order to solve the problem of local aliasing caused by insufficient image resolution and camera calibration error, three kinds of filters are designed: smooth filter, slice filter and normal filter. Effectively solve the target model on the drum, distortion and other problems that need to be optimized. Fifth, texture discontinuity problem usually occurs in image-based texture mapping because of the large difference in illumination. In this paper, a method of cutting the model by rotating tangent plane and generating synthetic texture is proposed. The texture is mapped to the surface of the target model. The effect is good and the problem of discontinuity of most textures is overcome. Finally, a three-dimensional human motion capture method based on analysis-synthesis is proposed to realize the human motion capture based on complex background video without marking points, which matches the model rendering image with the real image. The method of downhill simplex-simulated annealing is used to minimize the matching error. The experimental results show that this method can be used to capture motion accurately and stably.
【学位授予单位】：天津大学
【学位级别】：博士
【学位授予年份】：2009
【分类号】：TP391.41

【引证文献】