面向电力巡检机器人的SLAM算法研究与系统设计

发布时间：2018-05-03 23:17

本文选题：电力巡检机器人 + 角点特征　；参考：《浙江大学》2017年硕士论文

【摘要】：点云地图的创建算法是搭载2D Lidar的智能电力巡检机器人领域中一项关键性技术,点云地图的精度高低会直接影响到巡检机器人在工作过程中定位的精确度,进而影响到巡检机器人运动状态的更新以及路径规划的进行,是巡检机器人实现自主移动的根基所在,其重要性不言而喻。ICP算法是在创建点云地图过程中常用的一种算法,但是仅依靠ICP算法建立的点云地图随着地图建立时间的变长和地图覆盖范围的增大,其累积误差将会变得非常严重。闭环检测作为一种可以有效减小累积误差的手段,得到了国内外很多学者的广泛研究。闭环检测中的一个核心问题是地点识别,即能检测到在之前已经到过同一地点附近。解决地点识别问题的一种有效方法是提取单帧数据中的特征点,利用特征点来反映两帧数据之间的相似性。因此,如何设计针对2D Lidar的特征提取算法,以及如何利用提取出的特征来检索相似帧对于解决地点识别问题有着很明确的研究价值。因此本文第二章和第三章针对这两个问题展开。针对第一个问题,考虑到在实际环境中广泛存在的诸如建筑物墙角、桌角等稳定的角点特征,本文提出了一种基于2D Lidar的角点特征提取算法。算法结合两点间的欧式距离和相应法向量间的余弦距离双阈值来确定单帧点云中每点的邻域范围,具体而言,以较大的欧式距离阈值来确定粗略的邻域范围,再以较小的余弦距离来确定更加精准的邻域范围。同时为了更好地将角点从点云中提取出来,本文给出了一种新颖的评价函数,可以有效地检测出准确的角点。在网上公开的数据库上进行的对比实验显示本文所提出的角点特征提取算法的准确性较其他算法要更好。针对第二个问题,本文提出了基于2D Lidar角点特征的闭环算法。首先利用第二章中提出的针对2D Lidar的角点特征提取算法来获得单帧数据的签名,紧接着设计了一种相似帧判定方法让签名具有旋转不变性,同时给出了相似帧之间的相对位姿的计算方法,建立图模型,最后结合现有的图优化框架来对图模型进行后端优化。在网上公开数据库上的实验表明经过本文所提出的闭环算法优化后的点云地图相比未经优化的点云地图效果明显要更好。最后,针对与大立科技公司合作的电力巡检机器人建图及导航项目,本论文开发了一套结合建图、路径规划、实时导航功能的系统,并将所研究的相关算法应用到系统中,得到了很好的实用效果。目前该系统已经通过客户单位验收并交付使用。
[Abstract]:The algorithm of creating point cloud map is a key technology in the field of intelligent power inspection robot with 2D Lidar. The accuracy of point cloud map will directly affect the accuracy of location in the working process of the inspection robot. Furthermore, it affects the updating of the moving state and the path planning of the patrol robot. It is the foundation of the robot to realize the autonomous movement. The importance of ICP algorithm is self-evident. ICP algorithm is a common algorithm in the process of creating the point cloud map. However, the accumulated error of point cloud map based on ICP algorithm will become very serious with the increase of map establishment time and map coverage. Closed-loop detection, as an effective method to reduce the cumulative error, has been widely studied by many scholars at home and abroad. One of the key problems in closed-loop detection is location identification, which can detect that the location has been near the same location before. An effective method to solve the problem of location identification is to extract feature points from single frame data and use feature points to reflect the similarity between two frames of data. Therefore, how to design a feature extraction algorithm for 2D Lidar and how to use the extracted features to retrieve similar frames is of great value in solving the problem of location recognition. Therefore, the second and third chapters of this paper focus on these two problems. In view of the first problem, a corner feature extraction algorithm based on 2D Lidar is proposed in this paper, considering the stable corner features such as the corner of the building wall and the corner of the table, which are widely existed in the real environment. The algorithm combines the Euclidean distance between two points and the cosine distance between the corresponding normal vectors to determine the neighborhood range of each point in a single frame point cloud. In particular, a large Euclidean distance threshold is used to determine the rough neighborhood range. A smaller cosine distance is used to determine a more precise neighborhood range. At the same time, in order to extract the corner from the point cloud better, a novel evaluation function is given in this paper, which can detect the accurate corner effectively. A comparative experiment on a database published on the Internet shows that the proposed corner feature extraction algorithm is more accurate than other algorithms. To solve the second problem, a closed loop algorithm based on 2D Lidar corner feature is proposed. Firstly, the corner feature extraction algorithm for 2D Lidar is proposed in Chapter 2 to obtain the signature of single frame data, and then a similar frame decision method is designed to make the signature rotation-invariant. At the same time, the calculation method of the relative pose between similar frames is given, and the graph model is established. Finally, the back end of the graph model is optimized by combining the existing graph optimization framework. Experiments on the open database on the Internet show that the point cloud map optimized by the closed-loop algorithm proposed in this paper is more effective than the unoptimized point cloud map. Finally, in view of the power inspection robot mapping and navigation project of Dali Science and Technology Company, this paper develops a system which combines the functions of building map, path planning and real-time navigation, and applies the relevant algorithms to the system. Good practical results have been obtained. At present, the system has been accepted by the customer and delivered to use.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP242

【参考文献】