基于流形学习的多视角关节体运动跟踪与分析

项目名称： 基于流形学习的多视角关节体运动跟踪与分析

项目编号： No.60805012

项目类型： 青年科学基金项目

立项/批准年度： 2009

项目学科： 金属学与金属工艺

项目作者： 齐飞

作者单位： 西安电子科技大学

项目金额： 19万元

中文摘要： 目标跟踪是计算机视觉领域的重要研究内容，而关节体目标跟踪则是其中最具挑战性的问题之一。本课题围绕视觉目标跟踪问题，以关节体为核心研究对象，建立了关节体构形流形的数学描述与学习方法。本课题取得的主要研究成果有：在目标跟踪的滤波方法方面，提出了基于优化的粒子滤波方法与自适应窗宽的核粒子滤波方法；在特征选择方面，提出了基于Fisher线性判别的在线特征选择目标跟踪方法；在相关的目标检测与识别方面，分别提出了基于仿射曲率不变量目标识别方法和基于特征完整性的遮挡轮廓识别方法；在多视角处理方面，提出了基于稀疏表示的多摄像机协同跟踪优化策略；在相关的通用图像与视频处理方面，提出了基于稀疏的光流估计方法和基于自适应上下文模型的图像超分辨率恢复方法。 　本课题共获得国家发明专利授权3项，发表论文20篇，其中国际期刊1篇，国内期刊8篇，国际会议11篇；所有论文中，SCI检索4篇，EI检索14篇，ISTP检索4篇，很好的完成了预期研究目标。 　关节体运动分析是辅助训练、体感人机交互等应用领域的核心方法，本课题不仅为这些应用提供了理论与技术支撑，同时也为进一步深入分析关节体运动奠定了良好的基础。

中文关键词： 跟踪; 粒子滤波; 流形学习; 多视角

英文摘要： Objects tracking is an important field in computer vision, while articulated body tracking is one of its most challenging problems. In this project, we investigated the problem of tracking articulated bodies. We construct a mathematic model for the configuration manifold of articulated bodies and propose a method to learn the corresponding manifold. The achievements are in four aspects. On filtering methods, we improved particle filtering by providing an adaptive kernal bandwidth selection method and introducing an optimization procedure to obtain accurate ensity estimation. On mean shift based tracking, we propose a continuous linear projection determination method based on Fisher discriminant analysis for online feature selection. On object detect and recognition, we proposed affine invariant curvature and feature integrity based methods. On multiple camera collaboration, we proposed a sparse representation based convex optimization approach to get a real time adjust. For general image and video processing, we proposed a sparsity based optic flow estimation method and a context adaptive image super-resolution method. In this project, 3 patents have been filed and 20 papers have published. There are 1 paper published on an international journal, 8 on national journals, and 11 on international conferences. Our publications are with good quality because that, among them, there are 4 papers indexed by SCI, 14 by EI, and 4 by ISTP. To conclude, we have accomplished this project well. Articulated objects tracking is widely applicable in fields of sports training and advanced human computer interaction. The achivements of this project provide theory and technologies for these applications. In addition, we also laid a solid foundation for further investigation on more thorough analysis of the motion of articulated objects.

英文关键词： Tracking; Particle Filtering; Manifold Learning; Multiview