多视角下基于信息融合采样的3-D人体运动跟踪  

Multiview 3-D human motion tracking with information fusion sampling

作　　者：沈建锋[1] 杨文明[1] 廖庆敏[1] 

机构地区：[1]清华大学深圳研究生院,深圳518055

出　　处：《清华大学学报（自然科学版）》2012年第1期55-60,共6页Journal of Tsinghua University(Science and Technology)

摘　　要：以多视角同步视频为输入,提出了一个基于信息融合采样的3-D人体运动跟踪系统。系统先利用形状特征检测头部和躯干,然后实施单假设跟踪捕捉四肢。跟踪过程中,采样函数通过融合多种信息进行启发式采样,使粒子向受欢迎的方向跟踪,从而降低跟踪所需的粒子数。该系统结合检测和跟踪,有效地解决了漂移问题;通过信息融合采样提高了采样效率;在HumanEva-II数据库上的实验误差为71mm。与传统Condensation和退火粒子滤波方法相比,该系统在精度和采样效率上都有所提高。This paper presents a system for 3-D human motion tracking based on information fusion sampling from multiple synchronized videos.The system first detects the head and torso with shape features,and then performed a single hypothesis tracking to capture the limbs.The tracking fuses multiple kinds of information for heuristic sampling with the sampling function focused on the propagation of particles in plausible directions with a small number of particles.The combined detection and tracking alleviates drift and improves the sampling efficiency using information fusion sampling.The error on the HumanEva-II dataset was 71 mm.Thus,this method has better accuracy and sampling efficiency than the conventional condensation and annealed particle filtering methods.

关 键 词：模式识别 人体运动跟踪 采样函数 形状特征 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]