基于图优化的单目视觉SLAM技术研究

发布时间：2018-06-13 02:05

本文选题：机器人同步定位与建图 + 单目视觉　；参考：《华中科技大学》2016年硕士论文

【摘要】：机器人同步定位与建图(Simultaneous Localization and Mapping,简称SLAM)技术是机器人实现定位的主流研究方法。随着传感器和计算机视觉等技术的发展,视觉传感器越来越多的被装备到机器人上。单个摄像头具有性价比高,结构简单,适用范围广等优点,因此单目视觉SLAM方法受到越来越多的关注和研究。单目视觉SLAM方法主要分为基于滤波器的方法和基于图优化的方法,前者计算复杂度较高,无法满足实时性要求,因此,本文主要研究基于图优化的单目视觉SLAM方法。并行跟踪,局部建图和回环检测的单目视觉SLAM方法(ORB-SLAM)是目前比较完备和可靠的方法,本文对其进行了研究和实验,并针对其中的一些问题进行了改进。ORB特征检测算法只对一定范围内的旋转和尺度变化具有不变性,因此本文设计了评估实验对各种特征检测算法进行评测,找到了性能最佳的适合在单目视觉SLAM中使用的特征检测算法。特征的数目和分布对跟踪效果和相机姿态估计会产生影响,稳定的特征数目和均匀的特征分布会使算法更加鲁棒,本文设计了一种可以控制特征点提取的数目,并且使特征点在图像中尽量均匀分布的算法。在地图的初始化和局部建图中生成新的地图点时,ORB-SLAM使用的是线性三角化方法(Least Square Method,简称LS算法),该方法生成的地图点的位置有可能不准确,本文将其改进为迭代的三角化方法(Iterative Least Square Method,简称It-LS算法)。因为本文改进的特征检测算法使用的特征描述子为FREAK,因此将改进后的SLAM方法简称为FREAK-SLAM。针对以上改进,本文设计了相关的实验来进行验证。本文在室内、室外环境下对FREAK-SLAM进行了实验,实验表明FREAK-SLAM跟踪到的特征点的数目更加稳定,特征点的分布更加均匀,且相机的定位精度较ORB-SLAM有所提高,室内环境下平均提高了0.18cm,室外环境下平均提高了1.4m,提高的百分比约为10%。
[Abstract]:Simultaneous localization and mapping (slam) technology is the main research method for robot localization. With the development of sensors and computer vision, more and more vision sensors are equipped on robots. Single camera has many advantages, such as high cost performance, simple structure and wide range of application, so the single vision slam method has been paid more and more attention and research. Monocular vision slam method is mainly divided into filter based method and graph based optimization method. The former has high computational complexity and can not meet the real-time requirements. Therefore, this paper mainly studies Monocular vision slam method based on graph optimization. Monocular vision slam (Orb-slam) is a complete and reliable method for parallel tracking, local mapping and loop detection. The improved .Orb feature detection algorithm is only invariant to the rotation and scale change in a certain range, so this paper designs an evaluation experiment to evaluate the various feature detection algorithms. The best feature detection algorithm for monocular slam is found. The number and distribution of features will affect the tracking effect and camera attitude estimation, and the stable number of features and uniform feature distribution will make the algorithm more robust. In this paper, we design a method to control the number of feature points extracted. And the algorithm to make the feature points distribute evenly in the image as far as possible. When generating new map points in map initialization and local mapping, ORB-SLAM uses a linear triangulation method called least Square method (LS algorithm for short). The location of map points generated by this method may not be accurate. In this paper, the iterative triangulation method is improved as iterative least Square method (It-LS algorithm for short). Because the feature descriptor used in the improved feature detection algorithm is FREAK, the improved slam method is referred to as FREAK-SLAM. In view of the above improvements, this paper designed relevant experiments to verify. The experiment of FREAK-SLAM in indoor and outdoor environment shows that the number of feature points tracked by FREAK-SLAM is more stable, the distribution of feature points is more uniform, and the positioning accuracy of camera is higher than that of ORB-SLAM. The average increase was 0.18 cm in indoor environment and 1.4 m in outdoor environment. The percentage of increase was about 10 cm.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP242

【参考文献】