移动增强现实大范围定位与注册关键技术研究

发布时间：2018-03-09 07:42

本文选题：移动增强现实　切入点：大范围场景　出处：《华中科技大学》2013年博士论文　论文类型：学位论文

【摘要】：随着智能手机的广泛使用，移动AR技术更加受到国内外研究人员的关注。由于诸如智能手机之类的移动终端设备相对于PC机具有资源受限的特点，比如，计算速度慢、内存空间有限、手机功耗问题等，同时，智能手机轻便、体积小、可随身携带等特点，又扩大了使用者的活动范围，因此，需要在智能手机上提供能够完成大范围场景定位识别和三维注册的移动AR技术，但是，又不能简单的把应用于PC机上的增强现实技术移植到智能手机上。基于以上内容，本文提出了可以直接在移动设备上实现大范围场景的定位识别和三维注册的系统构架，主要研究工作如下：第一，通过使用重力来增强局部向量聚集描述符VLAD的鉴别力，设计了GAVLAD图像描述符，并设计了一个有效的向量量化策略，能将高维图像描述符压缩成几个字节的压缩编码，将图像描述符编码成几个字节，这样可以存储在移动设备中，有助于完成高效搜索，并以此设计了一个适合移动设备RAM的图像搜索引擎，使其能够高效完成移动设备上的定位识别。本文还建构了一个基于图像和传感器的相结合的压缩的索引结构，能在移动设备上直接处理大范围图像数据。第二，设计了一个简单有效的向量二值化方法以减少多特征融合内存占用量，并提出了一个位置敏感的融合算法将多特征融合起来。该算法可以将多个图像特征压缩成一个占用空间小的、高区分度的图像描述符，可以直接在移动设备上高效的进行存储和搜索。并提出将特征融合与索引结构联合优化，以提高定位识别的准确性，同时减少内存占用。第三，设计了一个灵活的摄像机初始化和追踪方法，可以以高达10Hz每帧的帧率在现阶段主流配置的手机上追踪非平面场景，一定程度上解决了大范围场景移动增强现实应用中的虚实注册问题。第四，发布了一套新的数据库，包含1,295,000个地理标记街景图像以及849个测试查询图像，这些资源可以被用作新的参照基准，可以在今后的相关研究中作为参照基准供其他研究人员继续使用。本文通过多组实验证明，，本研究提出的基于移动设备的大范围场景定位识别和三维注册移动AR系统提高了定位识别的精确性，并在节省内存、提高速度等方面取得了满意的效果，为提高移动增强现实系统的真实感和促进其走出实验室、面向广泛应用提供必要的技术支撑。
[Abstract]:With the wide use of smart phones, mobile AR technology has attracted more attention from researchers at home and abroad. Because mobile terminal devices such as smart phones have the characteristics of limited resources compared with PCs, for example, the computing speed is slow. Limited memory space, mobile phone power problems, and so on, at the same time, smart phones are light, small, portable and other characteristics, but also expand the range of activities of users, so, We need to provide mobile AR technology on the smartphone that can complete the large-scale scene location recognition and 3D registration, but we can't simply transplant the augmented reality technology applied to the PC to the smart phone. This paper presents a system framework that can directly realize the location recognition and 3D registration of large-scale scene on mobile devices. The main research work is as follows:. First, by using gravity to enhance the discriminant ability of the local vector aggregation descriptor (VLAD), the GAVLAD image descriptor is designed, and an effective vector quantization strategy is designed, which can compress the high-dimensional image descriptor into several bytes of compression coding. The image descriptor is encoded into several bytes, which can be stored on a mobile device, which is helpful for efficient search, and an image search engine suitable for the mobile device RAM is designed. This paper also constructs a compressed index structure based on image and sensor, which can directly process large range image data on mobile devices. Secondly, a simple and effective vector binarization method is designed to reduce the memory footprint of multi-feature fusion. A position sensitive fusion algorithm is proposed, which can compress multiple image features into a small space and high partition image descriptor. It can be stored and searched directly on mobile devices, and it is proposed to optimize the feature fusion and index structure to improve the accuracy of location identification and reduce the memory footprint. Thirdly, a flexible camera initialization and tracking method is designed, which can track non-planar scenes at a frame rate of up to 10 Hz per frame on the current mainstream mobile phone. To some extent, the problem of virtual reality registration in large scale mobile augmented reality applications is solved. In 4th, a new database containing 1, 295,000 geo-marked streetscape images and 849 test query images was released, which can be used as a new reference frame. It can be used as a reference for other researchers in future studies. In this paper, it is proved by many experiments that the large scale scene location recognition based on mobile device and 3D registered mobile AR system can improve the accuracy of location recognition and save memory. In order to improve the reality of mobile augmented reality system and promote it out of the laboratory, it provides the necessary technical support for the wide application of mobile augmented reality system.
【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：TP391.41

【参考文献】