【作者】 赵旭； 仝明磊； 刘允才；

【Author】 ZHAO Xu TONG Ming-lei LIU Yun-cai (Inst.of Image Processing & Pattern Recognition,Shanghai Jiao Tong University,Shanghai 200240)

【机构】 上海交通大学图像处理与模式识别研究所；

【摘要】 三维人体运动跟踪问题是一个复杂的高维非线性问题,本文针对三维人体运动重建的主要困难,通过对人体运动捕获数据的降维、聚类挖掘出人体运动的经验模型,在低维空间中以高斯混合模型表示人体运动参数的密度分布,然后与图像信息相结合,在生成式重建的框架内,以粒子滤波作为优化方法,重建出与序列图像对应的三维人体姿态。更多还原

【Abstract】 Tracking 3-D human motion is a complicated non-linear problem and suffering from curse of dimensionality. We present a novel approach to solve this difficult problem by digging the experience model from the motion capture data. The first step of the approach is reducing the dimensionality of the human pose space using predominant component analysis.And then,the k-means clustering is introduced to partition the iow-dimensionality space.The Gaussian Mixture Model plays a important role in the following step.The human pose parameter distribution is approximated by the GGM. Finally,in the frame of generative methods,we use the particle filter to reconstruct the 3-D human pose combining with image information.更多还原