CASIA OpenIR  > 毕业生  > 博士学位论文
基于局部统计特征的手势分析
其他题名Local Statistic Feature Based Gesture Analysis
陈宇峰
学位类型工学博士
导师王阳生
2005-07-01
学位授予单位中国科学院研究生院
学位授予地点中国科学院自动化研究所
学位专业模式识别与智能系统
关键词统计形态分析 局部矩分析 手势识别 人机交互 Statistic Shape Analysis Local Moment Analysis Gesture Recognition Interactive Interface
摘要随着计算机技术的发展,传统的人机交互技术已难以适应越来越复杂多样的需求。用户要求更加自然和智能的交互方法,包括声音、视觉和智能传感器等等。其中基于计算机视觉的方法具有自然、方便的特点,受到了广泛的关注。其中手势更是其中最具有自然交互性的方式,目前在计算机游戏中得到了很好的应用。 根据游戏中实时性要求高的特点,不同于一些基于模型的方法和整体分析方法,我们采用了基于局部统计特征的方法,尽量简化运算复杂度,以实现实时控制的目的。另一方面,为了满足游戏控制指令自然多样的要求,我们在基本特征的基础上,对特征结构进行详细的分析,以识别不同的手势。本文从简单的轮廓分析与识别,到图像特征的提取与结构分析,进行了多种方案的分析比较,并进行了总结。所研究的主要内容如下: 首先,基于“曲率的分布可以完全反映曲线的基本特征”的思想,从简化的轮廓曲线入手,在微分曲线的基础上,提出了一种基于曲率分布统计分析的曲线特征提取的方法。并结合二维信息,对区域曲线段的模型建立,以及相应的检测算法进行了研究,实验表明基于曲率的分析,具有高效准确的特点。 基于“局部特征集合可以高效表达整体结构”的想法,在轮廓特征分析的基础上,我们提出了基于曲率墒的曲线结构分析方法,并对分析结果进行了结构化建模。然后通过一种投票策略,优化模型匹配的寻优过程,以实现快速的形状识别的目的,并取得了良好的实验效果。 基于“基于矩的统计量可以高效并完全表示图像分布信息”的思想,我们提出了一种具有几何,线性光照,对比度不变性的矩分析方法,并给出其快速算法,实验结果表明,此方法对特征检测具有很好的区分度,结果令人满意。 基于“多种信息的结合可以增强系统性能的”思想,我们尝试了使用肤色、运动、频率等多种信息,从立体视觉获得更高的自由度,用结构分析方法得出更多的信息量,并将这些信息有机的结合起来,实现了一系列较为成功的应用。
其他摘要With the development of computer sciences, the traditional Human Machine Interface (HMI) becomes limited for the more and more complicated equirements. The natural and intelligent interactive methods are reported on the literature, such as voice, vision and advanced sensor based methods. The computer vision based method is getting more and more attention for its nature and convenience. And the gesture is one of the most interactive means that has been applied in many computer games recently. To meet the requirements of real-time game control, a local statistic based method is adopted to detect features, in stead of the model based methods or global analysis methods. By this means the computation of the feature is sim plified and thus the high effciency is obtained. On the other hand, based on the basic local features, the structural character of feature group is analyzed to get more meanings from the complicated gestures for more complicated games. From the simple contour to the image, and from the feature to the group, we tried many methods for analyze and comparison. Focusing on this topic, this dissertation mainly includes the following research issues: Firstly, basing on the idea that the curve can be totally expressed by the distribution of its curvature, we proposed a novel differential contour based method to analyze the distribution of the curvature. Combined with some of the two dimensional information, the local curve curvature model are founded and related detection method is carried out. The experiment shows the curvature based method is effcient for the real time requirement. Secondly, based on the fact that the structure of the global model can be described effciently by the local feature groups, a curvature entropy based structure analysis method is introduced to get the structural model of certain shape. Then by a voting scheme, the shape is effciently matched with the models. The experiment shows it suitable for shape comparison. Then based on the notion that moment is the complete and effcient statistic of the image, the novel geometry, lighting and contract invariant moment is proposed as well as its fast algorithm. Experiment shows the detection results of the method are differentiable and satisfiable. At last, to fully implement the idea that the system can be promoted by the multiple cues, we tried many methods include structure information, skin, motion, and frequency information. The more degrees of freedom are obtained from the stereo vision cue, and the global shapes are compared with the structure cue. With all the cues confused together some vision based games are successfully developed.
馆藏号XWLW0
其他标识符200218014603195
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/5876
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
陈宇峰. 基于局部统计特征的手势分析[D]. 中国科学院自动化研究所. 中国科学院研究生院,2005.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
陈宇峰200218014603195.p(34644KB) 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[陈宇峰]的文章
百度学术
百度学术中相似的文章
[陈宇峰]的文章
必应学术
必应学术中相似的文章
[陈宇峰]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。