基于多模型融合的状态预测算法研究.docVIP

基于多模型融合的自然人手跟踪研究? 林燕，冯志全，朱德良，尚爱丽 济南大学信息科学与工程学院，济南，250022 山东省网络环境智能计算技术重点实验室，济南，250022 Multi-Model Fusion LIN Yan, FENG Zhi-quan, ZHU De-liang, SHANG Ai-li School of Information Science and Engineering, University of Jinan, Jinan, 250022, China Shandong Provincial Key Laboratory of Network based Intelligent Computing, Jinan, 250022, China Abstract：In this paper, we propose an algorithm based on multi-model fusion in the state prediction of particle filtering to improve the precision of 3D hand tracking. First, through the analyzing the experimental data of the virtual oven system, combining with the behavior understanding and description of hand, we construct a model of human hand behavior based on cognitive psychology. Second, constructing the hand motion model by the data of hand tracking process, and using Sigma points to the motion model data, we can get the local analyzing prediction model. Finally, we fuse the two model into a model according to the similarity with the current frame image, which is the state prediction model in the process of particle filter. The innovation of this paper are: (1) establishing two different models of hand motion; (2) the fusion of the two models according to the similarity with the current frame image. Compared with annealing algorithm in the same running time, our algorithm improves the precision of 3D hand tracking. Keywords: particle filter; cognitive-behavioral model; state prediction model; Sigma points;3D freehand tracking 引言 人手跟踪是智能人机交互的关键技术之一，也是国内外计算机视觉领域的研究热点。精确的人手跟踪可以促进人机交互、手语跟踪识别、视觉监控、机器人控制等许多应用的发展[1~2]。如何获得精确地人手跟踪也是计算机视觉和人机交互领域中一个富有挑战性的重要问题。2009年，Robert Y. Wang 等人在其SIGGRAPH论文中[3]，让操作者带上彩色手套，他们对彩色区域块进行编号，建立手势数据库及其采样搜索机制，实现对手势的三维实时跟踪。Andreas Aristidou [4]利用光学运动跟踪系统3D人手跟踪，但要求在手的关键部位打上标记。这种戴数据手套跟踪或者是贴标记来跟踪人手的方法，是借助一些辅助设备来实现三维人手跟踪，不满足人手自然跟踪的特性。 自然人手的三维运动跟踪问题是一个典型的非线性状态估计问题, 目前跟踪三维人手的方法主要分为两大类：基于表象的方法[5-6]和基于模型的方法[7-9]。基于表象的跟踪方法本质在于建立图像特征空间到状态空间（人手空间）的映射关系，利用图像特征直接得到手势状态。这类方法容易满足跟踪系统的实时性要求，但人手运动过程中人手表象的剧烈变化致使该空间映射必定是高度非线性的