【作者】 温正东；

【导师】 闵华清；

【作者基本信息】 华南理工大学 ， 计算机应用技术， 2011， 硕士

【摘要】 随着社会的进步和科学技术的快速发展,机器人技术的应用越来越广泛,在众多的科学领域、工业部门和家庭生活中发挥着重要的作用,而人与机器人需要交互的场景也越来越多,为了提高机器人与人的交互性和协作性,需要对人—机器人交互系统进行研究。在当前的研究中,基于视觉的手势识别因为摄像机的低廉成本和手势的直观性,通用性和强大语义使得其在人机交互研究中具有强大的潜力。在目前的基于视觉的手势识别研究中,依靠肤色信息来建立模型进而识别手势是较为高效实用的方法,但是由于计算机处理数字摄像机采集到的肤色信息极易受到光照和背景变化的影响,现有的手势识别系统大多在室内固定环境下才能取得理想的识别效果,而在移动的复杂环境下不具有工作能力。针对这一问题,本文主要研究用于移动机器人的基于肤色的手势识别系统,研究的重点集中在肤色模型的建立,复杂环境下的肤色分类和模型的自适应问题。本文在建立单高斯肤色模型的基础上,提出了一种基于前景加权直方图的动态阈值调整算法,该算法通过肤色模型得到肤色概率图后,在目标区域被识别的前提下,对目标区域的像素的概率建立加权直方图,通过分析该直方图调整肤色分类阈值,在一定程度上解决光照变化对肤色的影响问题。针对模型对光照和背景变化的自适应问题,文章提出了静态单高斯模型与动态直方图方法结合的自适应算法,该算法以阈值调整和肤色分割结果为依据,在满足一定条件的情况下通过建立前景与背景直方图对单高斯模型进行调整,具有很好的实时性,同时对光照和背景的变化有很好的适应性。文章最终将人手分割,轨迹跟踪,模型调整,手势分析等手势识别的关键环节串联起来,建立起一个可用于移动智能机器人人机交互中的手势识别系统,实验结果表明该系统准确高效,具有实用价值,并且实验方法具有通用性,可被扩展到各种手势识别交互系统中。更多还原

【Abstract】 With the fast development of technology, robot has been broadly applied, and has played an important role in many scientific fields as well as some industrial sections and family sections, so human need to interact with robots more. In order to improve the interaction and collaboration between human and robot, Human—Robot Interactive System should be researched. In the current study, vision based hand gesture recognition has strong potential because of its intuitivism, versatileness, powerful semantic and camera’s low cost.In current research on vision based hand gesture recognition, building skin model to recognize hand gesture is an efficient method. However, skin information captured from digital camera can be effected by lighting condition and background varying. Most current hand gesture recognition systems get good result under indoor and static environment but not mobile complex environment. In order to solve this problem, this paper mainly research hand recognition system based on skin for mobile robots. It focuses on building skin model, skin classification under complicated environment and the adaptability of skin model.After building a Single Gaussian Skin Model, we propose a dynamic threshold adjust algorithm based on foreground weighted histogram. The algorithm gets skin probability through Gaussian Model and segments object area, then builds weighted histogram of the area. According to analyzing the histogram, the threshold can be adjusted, which to some extent solves light condition varying problem.This paper proposes an adaptive model algorithm which combines static single Gaussian model and dynamic histogram model for solving lighting and background problem. According to the result of threshold adjusting and image segmentation, the algorithm adjusts the single Gaussian model through building foreground and background histogram. This method is fast and can adapt to lighting and background varying.We finally integrate hand segmentation, trajectory tracking, model adjustment and gesture analyzing into a gesture recognition system for human—robot interaction. Experiment results show that the system is accurate and efficient. And the method is general and extendable, can be used in other hand gesture recognition interaction system.更多还原