基于RGB-D的SLAM方法改进研究

发布时间：2018-11-06 16:19

【摘要】：SLAM(Simultaneous Localization and Mapping,同时定位与地图创建)是公认的实现机器人自主导航的核心环节,也是目前最具挑战性的课题之一。以Kinect为代表的RGB-D传感器,不仅能采集周围环境的彩色信息还能直接获取其对应的深度信息,数据处理过程简单,适合用于三维地图重建。基于RGB-D传感器进行的SLAM研究被称为RGB-D SLAM,是目前机器人自主导航领域较为火热的一个研究课题。为了解决原有RGB-D SLAM方法中存在的效率低和误差大等问题,本文分别对其流程的前端和后端进行了研究及改进,得到一种准确性、鲁棒性和实时性较好的RGB-D SLAM方法。其具体研究成果如下:首先,研究了用于采集周围环境中RGB-D信息的Kinect传感器的工作原理、内外参模型以及标定方法。用Matlab里的联合标定工具箱,对Kinect的彩色镜头和深度镜头进行标定对齐,并将标定对齐前后得到的点云图片进行比较分析,以验证标定有助于提高RGB图片像素点和Depth图片像素点的正确匹配率。其次,基于对RGB-D SLAM流程前端的几个环节,即特征检测与描述子提取、特征匹配、错误匹配剔除、运动变换估计和运动变换优化的研究,提出一种将双相匹配法与阈值法相结合的改进的错误匹配剔除算法,该算法用时更少(用于SIFT、SURF和ORB算法上,分别减少了14.3%、14.7%和58.6%),同时保留的正确匹配点数目更多(用于SIFT、SURF和ORB算法上,分别增加了5.7%、34.7%和26.9%)。然后,基于对RGB-D SLAM方法后端的几个环节,即位姿图的生成、闭环检测、位姿图的优化、运动轨迹和三维点云地图生成的研究,提出一种将近距离逐帧闭环检测、远距离随机闭环检测以及BoVW的思想相结合的改进闭环检测算法,基于该算法生成的位姿图更为整洁且消耗的时间更少。最后,采用公开数据集Computer Vision Group和相应的结果评估工具对改进前后的RGB-D SLAM方法进行了评估,验证了该改进RGB-D SLAM系统在构建地图的准确性和实时性上均有所提高。另外,用Turtlebot机器人搭载Kinect进行了场地实验,该系统可以在机器人运行的过程中生成(不断更新)位姿图和三维点云地图,验证了该改进RGB-D SLAM方法的鲁棒性和有效性。
[Abstract]:SLAM (Simultaneous Localization and Mapping, simultaneous location and map creation) is recognized as the core of autonomous navigation for robots, and it is also one of the most challenging topics at present. The RGB-D sensor represented by Kinect can not only collect the color information of the surrounding environment but also obtain the corresponding depth information directly. The process of data processing is simple and suitable for 3D map reconstruction. The research of SLAM based on RGB-D sensor is called RGB-D SLAM, which is a hot research topic in the field of robot autonomous navigation. In order to solve the problems of low efficiency and large error in the original RGB-D SLAM method, the front-end and back-end of the process are studied and improved in this paper, and an accurate, robust and real-time RGB-D SLAM method is obtained. The specific research results are as follows: firstly, the working principle, internal and external parameter model and calibration method of Kinect sensor used to collect RGB-D information in the surrounding environment are studied. Using the joint calibration toolbox in Matlab, the color lens and depth lens of Kinect are calibrated, and the point cloud images before and after calibration are compared and analyzed. Verification and calibration are helpful to improve the correct matching rate between RGB image pixels and Depth image pixels. Secondly, based on the research of several links in the front end of RGB-D SLAM process, such as feature detection and descriptor extraction, feature matching, error matching and culling, motion transform estimation and motion transformation optimization. This paper presents an improved error matching elimination algorithm which combines the biphasic matching method with the threshold method. The algorithm takes less time (14.37% for SIFT,SURF and 58.6% for ORB). At the same time, the number of correct matching points retained is more (for SIFT,SURF and ORB algorithms, an increase of 5. 7% and 26. 9% respectively). Then, based on the research of several links in the back-end of RGB-D SLAM method, the generation of the spot pose graph, the closed-loop detection, the optimization of the pose map, the motion path and the 3D point cloud map generation, a close-loop detection based on the close distance frame by frame is proposed. Based on the improved closed loop detection algorithm based on the combination of long distance random closed loop detection and BoVW, the position and pose graph generated by the algorithm is cleaner and consumes less time. Finally, the RGB-D SLAM method before and after the improvement is evaluated by using the open data set Computer Vision Group and the corresponding result evaluation tools, and the accuracy and real-time performance of the improved RGB-D SLAM system are proved to be improved. In addition, the Turtlebot robot is used to carry on the field experiment with Kinect. The system can generate (update) the pose map and 3D point cloud map while the robot is running. The robustness and effectiveness of the improved RGB-D SLAM method are verified.
【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】：TP242

【参考文献】