基于栅格粒子滤波的人体头肩三维姿态估计

为了实现单目视频中人体头部和肩部位置的自动跟踪,并估算出人体头肩部的三维姿态.提出了一种新的基于头-肩三维模型的栅格粒子滤波方法,使用Chamfer距离进行边缘相似性度量.算法使用一个预先生成的离散化的栅格状态空间,在每一个时间步将该状态空间的一个较小的子空间作为粒子的搜索空间,有效地限制了所用粒子数量,提高了粒子滤波过程的执行效率.由于三维信息的缺失,单目视频中人体三维姿态估计一直是一个难点问题.而实验结果表明,本方法能够有效并准确地估计出人体头肩部的三维姿态,为人体姿态分析和行为理解的研究提供了一个新的基础.

Grid particle filter based monocular 3D head-shoulder pose estimation

In order to track and estimate 3D head-shoulder poses in monocular videos,we propose a new 3D head-shoulder model based particle filter.Edge cue and chamfer distance are used to build the likelihood function of the proposed particle filter.In each time step,the proposed algorithm generates a discretized grid state space and takes a subspace of it as the searching space of particles.Therefore,the number of particles is well limited and the performance of the particle filter is improved.The experimental results prove that our method is effective and robust for the tracking and estimation of 3D head-shoulder poses.The study provides a new basis for the human pose analysis and behavior understanding.