一种基于改进UKF的3D人手跟踪算法  

A Novel 3D Moving Human Hand Tracking Algorithm Based upon Improved Unscented Kalman Filter

作　　者：冯志全[1] 孟祥旭[1] 

机构地区：[1]山东大学计算机科学与技术学院

出　　处：《计算机辅助设计与图形学学报》2006年第8期1264-1269,共6页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(60473103;60473127);国家"八六三"高技术研究发展计划(2003AA414310)

摘　　要：将手指作为基本处理对象,对UKF(unscented Kalman filter)算法进行改进,并利用它对当前手指各关节进行预测;以预测值作为初值,用局部搜索技术对误差较大的关节用改进的UKF算法重新进行预测,直到该手指在像平面上的投影轮廓和图像轮廓之间的距离图满足指定的精度为止.该算法以状态变量量测值的获取作为突破口,解决现有算法中跟踪精度过分依赖于3D人手模型精度的问题.实验结果表明,该算法具有较强的处理局部自遮挡问题能力,对3D人手模型的不精确性也具有更好的鲁棒性.A novel moving 3D human hand tracking algorithm based on unscented Kalman filter（UKF） is put forward in this paper. In our algorithm, a finger is regarded as a process unit within each loop. Initially, the algorithm predicts states of all joints of a finger using improved UKF algorithm of this paper. The predicted states are used as initial states of a loop process to the finger. Then, re-predicts those joints using the improved UKF algorithm again and again until the differences of all corresponding features between the silhouettes of the image and the projection of the 3D predicted virtual hand satisfy the given precision. Existing algorithms have a problem： the tracking precision relies heavily on 3D human model. By obtaining the observations of the states, our algorithm resolves the problem. Compared with the relevant algorithms, our algorithm can more effectively deal with hand self-occlusion issues, and is more robust to 3D human hand models as well.

关 键 词：卡尔曼滤波 无味卡尔曼滤波 运动人手跟踪 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]