【作者】 杨阳；

【导师】 史彩成；

【作者基本信息】 北京理工大学 ， 生物医学工程（专业学位）， 2016， 硕士

【摘要】 增强现实(Augmented Reality,简称AR)是一门新兴发展的虚实融合技术,主要通过对真实场景叠加虚拟物体实现图像信息增强,目前已在军事、医疗、工业和娱乐等领域展开深入研究。解决增强现实几何一致性问题需要系统各部分相互配合,例如摄像机跟踪、目标三维注册、几何场景重建等。图像特征点是连接增强现实各个流程的桥梁,其算法的鲁棒性和实时性直接关系到增强现实系统的性能。因此,研究与优化特征点相关算法成为解决增强现实几何一致性问题的关键。图像特征点可以通过摄像机成像原理应用于摄像机跟踪,特征点的提取、匹配和跟踪是其基本操作。通过研究,发现Harris、SUSAN和FAST等算法无法满足增强现实对系统鲁棒性的要求;SIFT和SURF算法无法满足增强现实对系统实时性的要求;本文以SURF算法为基础,根据E-FAST算法在实时性方面的优势,对SURF算法加速优化。在特征点匹配研究中,本文以欧几里德距离为相似性度量准则,使用kd-tree进行数据索引,并通过RANSAC算法剔除误匹配特征点;在特征点跟踪方面,本文以KLT光流算法为基础,引入局部自适应性对算法进行优化;针对以上算法研究和优化,本文均进行理论分析与实验对比。图像特征点相关算法均涉及到大量浮点运算和卷积运算,基于CPU的硬件平台难以保证增强现实系统的实时性。利用GPU和DSP处理器在图像处理方面的优势和并行化结构,在像素和特征点级别对SURF和KLT算法进行并行化加速,实现算法优化的目的。最后,使用ARToolKit增强现实工具实现具体例程开发。更多还原

【Abstract】 Augmented Reality is a newly developed fusion actual situation,mainly through a virtual object superimposed on a real scene information of image enhancement,has in-depth research in the military,medical,industrial and entertainment fields.Augmented Reality solve geometric consistency of the various parts of the system need to cooperate with each other,such as camera tracking,3D registration target,geometric scene reconstruction.The image feature point is to connect the various processes of the bridge augmented reality,robustness and real-time directly related to their algorithms to enhance the performance of real systems.Therefore,the study and optimization of feature points algorithm becomes the key to solving the geometric consistency of augmented reality.It can be applied to the image feature point camera tracking camera imaging principle by extracting feature points,matching and tracking of its basic operations.Through research,we found Harris,SUSAN and FAST algorithms can not meet the augmented reality system robustness requirements;SIFT and SURF algorithm augmented reality system can not meet the real-time requirements;paper SURF algorithm,according to E-FAST algorithm advantages in terms of real time,on the SURF algorithm acceleration optimization.In the feature point matching study,this paper Euclidean distance as the similarity measure,use kd-tree index data,and through RANSAC algorithm eliminate false matching feature points;the feature point tracking,optical flow algorithm paper KLT based on the introduction of local adaptive algorithm optimization;algorithm and optimization for the above article are theoretical analysis and experimental comparison.The image feature points related algorithms involve a large number of floating-point operations and convolution-based hardware platform CPU is difficult to ensure enhanced real-time reality system.The use of GPU and DSP processor in image processing advantages and parallel structures,and feature points in the pixel level for SURF and KLT algorithm parallelization accelerated algorithm optimization purposes.Finally,the use of augmented reality tools to achieve specific ARToolKit routine development.更多还原