基于视觉的机器人室内定位算法研究与嵌入式实现

发布时间：2018-02-09 23:39

本文关键词： 移动机器人嵌入式系统 SLAM 计算机视觉 CUDA　出处：《哈尔滨工业大学》2017年硕士论文　论文类型：学位论文

【摘要】：解决机器人在未知环境中的定位问题,是实现机器人智能化的关键。针对消费级室内移动机器人的设计需求,为了使其利用机载传感器在未知的室内环境中实现定位并降低制造成本,本文讨论了如何利用视觉传感器与嵌入式处理器,实现基于视觉的同步地图构建与定位算法。本文首先介绍了视觉SLAM算法的原理,将其分为前端方案与后端方案两部分进行可行性讨论。在前端方案部分,讨论了特征法与直接法视觉里程计的优缺点,通过实验比较了各种主流特征点的运算速度与匹配性能,选择了双目摄像头作为视觉传感器,并以基于ORB的特征法视觉里程计作为前端方案;在后端方案部分,分析了回环检测的必要性,讨论了以EKF为代表的滤波法和以图优化法为代表的优化法,选择了图优化法作为后端方案。根据方案论证,本文选择了ORB SLAM算法作为嵌入式视觉SLAM的实现方案,分析了ORB SLAM算法流程,并将其移植到Nvidia Tegra K1嵌入式板卡中。针对算法在嵌入式系统中运算效率较差、计算精度较低的问题,对代码进行了优化与参数调整,并利用基于CUDA的GPU加速技术对ORB特征的提取与匹配进行加速,提高了算法在嵌入式系统中的运行效果,增加了该系统的实用性。为了评估定位算法的实时性和有效性,在嵌入式视觉SLAM算法中进行EuRoC双目视觉数据集测试。通过对输出轨迹的定量分析,证明了算法移植与优化的有效性。最后将双目摄像头与嵌入式板卡组合成完整的室内定位系统,利用磁导轨车进行室内环境下的场地测试实验,证明了基于视觉的嵌入式室内定位系统的有效性。
[Abstract]:To solve the problem of robot location in unknown environment is the key to realize robot intelligence. In order to use airborne sensors to locate in unknown indoor environment and reduce manufacturing cost, this paper discusses how to use vision sensors and embedded processors. This paper first introduces the principle of visual SLAM algorithm, and divides it into two parts: front-end scheme and back-end scheme for feasibility discussion. The advantages and disadvantages of the feature method and the direct method are discussed. The computing speed and matching performance of various mainstream feature points are compared through experiments. The binocular camera is chosen as the vision sensor. In the back end part, the necessity of loop back detection is analyzed, and the filtering method represented by EKF and the optimization method represented by graph optimization method are discussed. This paper chooses the graph optimization method as the back-end scheme. According to the scheme demonstration, this paper chooses the ORB SLAM algorithm as the implementation scheme of embedded visual SLAM, and analyzes the ORB SLAM algorithm flow. The algorithm is transplanted to the Nvidia Tegra K1 embedded board. The code is optimized and the parameters are adjusted in order to solve the problem of poor operation efficiency and low accuracy of the algorithm in the embedded system. GPU acceleration technology based on CUDA is used to accelerate the extraction and matching of ORB features, which improves the running effect of the algorithm in embedded system and increases the practicability of the system. The EuRoC binocular visual data set is tested in embedded visual SLAM algorithm. Finally, the binocular camera and the embedded board are combined into a complete indoor positioning system, and the magnetic guide rail vehicle is used to carry out the field test experiment in the indoor environment. The validity of embedded indoor positioning system based on vision is proved.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【参考文献】