基于Kinect的人手姿态混合跟踪方法  

Hybrid Method for Hand Articulations Tracking Using Kinect

作　　者：乔体洲[1] 戴树岭[1] 

机构地区：[1]北京航空航天大学自动化科学与电气工程学院,北京100191

出　　处：《计算机辅助设计与图形学学报》2015年第4期713-720,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家"八六三"高技术研究发展计划(2009AA01Z333)

摘　　要：为了实现在无标记的情况下不对称地跟踪人手及其所有关节的位置和姿态,提出一种软硬件结合的混合跟踪计算框架,同时使用电磁跟踪器和无标记人手关节姿态分析算法提出基于CUDA的异步并行粒子群优化(PSO)加速方法.首先通过跟踪器测量人手手腕的位置姿态,使用Kinect数据作为输入,在三空间(双颜色空间和深度空间)下进行手部区域分割;然后使用PSO方法将手关节的23个自由度的跟踪问题转化为求解一个优化问题,使用不对称策略来提高部分手指的跟踪性能,寻找给定参数空间内能够最小化观测值和估计值之间偏差的手模型参数解.该方法不需要进行任何标记,可以对手部关节姿态进行连续跟踪,实验结果表明,其在实验的硬件平台上可以达到12帧/s的运行速度,平均误差稳定在10 mm以内.In order to track position and orientation of hand and its articulations, we implement a hybrid solution for markerless hand articulation tracking with electromagnetic tracker and visual observations from Kinect. The position tracking of hand uses an electromagnetic tracker, and tracking of 23 DOFs of articulations is treated as an optimization problem which is solved using Particle Swarm Optimization (PSO), and it is essentially the progress of seeking for the hand model parameters that minimize the discrepancy between the hypothesized hand model and actual hand observations. The system input is RGB and depth images from Kinect sensor. This system does not require special markers and is capable to track hand articulations frame by frame. Experiments show that ro-bust 3D tracking of hand articulations can be achieved in 12FPS, and average error is less than 10mm.

关 键 词：手关节跟踪 无标记跟踪 粒子群优化 电磁跟踪器 GPU加速 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]