基于图像检索的定位算法研究与实现

发布时间：2018-05-11 08:28

本文选题：图像检索 + 定位算法　；参考：《山东大学》2017年硕士论文

【摘要】：基于图像的定位算法作为一门融合计算机视觉、机器学习、多视图几何、图像检索等众多科研领域的交叉性学科技术,在机器人导航定位、现实增强、三维重建、地标识别等领域有着关阔的应用前景和巨大的研究价值。基于图像的定位技术也是高层次图像处理任务不可缺少的组成部分,其典型应用包括图像中物体的语义标注、通过提供相机初始姿态加速大规模SFM模型重建等。本文在阐述相机模型、PNP算法、RANSAC算法、SFM算法、CNN理论、传统VLAD算法的基础上,总结归纳了基于图像检索、基于特征匹配和基于逻辑回归三大主流图像定位算法的优缺点,并提出了基于VLAD-CNN图像检索的定位算法。本文的研究内容和主要工作如下:(1)在对现有算法和理论深入理解的基础上,实现了 SPTR-RANSAC算法、P4Pf算法。从包含噪声数据的2D-3D匹配关系中,利用融合P4Pf的SPTR-RANSAC算法可快速、有效的求解待定位图像的6-DOF。(2)针对图像检索过程中,实时构建k-d树搜索结构容易造成算法效率低下的问题,本文实现了 GPU版本的KNN(K Nearest Neighborhood)算法,用于加速邻近点的查找。通过GPU强大的并行执行能力减少特征匹配所需时间,提高基于图像检索定位算法的效率。(3)针对传统图像检索定位算法的不足,本文提出了基于VLAD-CNN的图像定位算法,并详细论述了 NetVLAD网络层基本原理,介绍了 VLAD-CNN网络的基本结构、基于SFM模型的自动样本标定方法和训练过程中采用的损失函数。通过与现有主流算法在公共数据库上的对比实验,验证基于VLAD-CNN图像检索定位算法的合理性和有效性。
[Abstract]:The image-based localization algorithm is a cross-disciplinary technology that integrates computer vision, machine learning, multi-view geometry, image retrieval and so on, in robot navigation, reality enhancement, 3D reconstruction, and so on. Landmark recognition and other fields have the application prospect and great research value of Guan Kuo. Image-based localization is also an indispensable part of high-level image processing. Its typical applications include semantic tagging of objects in images and acceleration of large scale SFM model reconstruction by providing camera initial attitude. In this paper, the advantages and disadvantages of three main image localization algorithms based on image retrieval, feature matching and logical regression are summarized on the basis of describing the camera model VLAD algorithm and the traditional VLAD algorithm. A localization algorithm based on VLAD-CNN image retrieval is proposed. The research contents and main work of this paper are as follows: 1) on the basis of deep understanding of the existing algorithms and theories, we implement the SPTR-RANSAC algorithm and P4Pf algorithm. From the 2D-3D matching relationship of noisy data, using the SPTR-RANSAC algorithm of fusion P4Pf can solve the 6-DOF.F-2 of the image to be located quickly and effectively. In the process of image retrieval, constructing the search structure of k-d tree in real time will lead to the problem of low efficiency of the algorithm. In this paper, the GPU version of KNN(K Nearest neighbor algorithm is implemented, which is used to speed up the search of adjacent points. Through the powerful parallel execution ability of GPU to reduce the time required for feature matching and to improve the efficiency of image retrieval localization algorithm, aiming at the shortcomings of traditional image retrieval and localization algorithms, this paper proposes an image location algorithm based on VLAD-CNN. The basic principle of NetVLAD network layer is discussed in detail. The basic structure of VLAD-CNN network, the automatic sample calibration method based on SFM model and the loss function used in the training process are introduced. The validity and rationality of the algorithm based on VLAD-CNN image retrieval are verified by comparing it with the existing mainstream algorithms in the common database.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】