基于旋转二维激光的三维地图构建
发布时间：2018-07-16 22:03
【摘要】：当下很多机器人需要在工作场景中移动完成任务,而地图是移动机器人理解周围环境并作出行动的基础和关键。本文叙述了一种基于2D激光的环境三维地图构建方法,使机器人系统可以实现在动态运动过程中采集激光数据并构建三维点云地图。本文针对移动地图构建,以通过二维旋转激光传感器构建三维点云地图为目的,在系统架构、传感器数据处理、位姿估计等方面展开研究。主要研究内容和成果如下:(1)设计了环境感知单元的软硬件架构,并针对多传感器原始数据时间戳不一致问题,提出了基于样条平滑的多传感器时间同步。对不同的传感器采用了不同的时间同步方法,一般传感器数据直接使用线性插值法,而对于存在时间不确定性的视觉里程计数据,则利用重投影误差和关键帧信息,采用样条平滑方法实现时间同步。(2)构建了 JDL 模型(Joint Directors of Laboratories Model)和 EKF(Extended Kalman Filter)相结合的多传感器融合位姿估计方法。该方法采用JDL模型作为多传感器融合架构,在JDL模型的态势推断层级,采用EKF方法融合基于ORB-SLAM2的双目视觉里程信息和惯性传感单元数据。通过两者在位置变换估计和角度变换估计的优势互补提高了位姿估计精度。(3)实现了基于ICP(Iterative Closest Point)方法的在线和离线多帧地图融合优化。在线模式下,本文将多个点云局部地图使用ICP的方法拼接融合在一起,得到最终的激光地图。在保证基本的实时性的同时,最大可能地消除了实时姿态估计的累计误差。离线模式下,基于在线模式的基础,添加了基于ICP方法的单个局部点云内部优化。
[Abstract]:Nowadays, many robots need to move in the work environment to complete the task, and map is the basis and key for the mobile robot to understand the surrounding environment and take action. In this paper, a 3D map construction method based on 2D laser is described, which enables the robot system to collect laser data and construct 3D point cloud map in the course of dynamic motion. In order to construct 3D point cloud map by two-dimensional rotating laser sensor, this paper studies the system architecture, sensor data processing, position and pose estimation and so on. The main research contents and results are as follows: (1) the hardware and software architecture of the environment sensing unit is designed, and the multi-sensor time synchronization based on spline smoothing is proposed to solve the problem of multi-sensor original data timestamp inconsistency. Different time synchronization methods are used for different sensors. Generally, linear interpolation is used directly for sensor data, while for visual odometer data with time uncertainty, reprojection error and key frame information are used. The spline smoothing method is used to realize time synchronization. (2) A multi-sensor fusion position and attitude estimation method based on JDL (Joint Directors of Laboratories Model) and extended Kalman filter (EKF) is proposed. In this method, the JDL model is used as the multi-sensor fusion framework, and the binocular visual mileage information and the inertial sensor unit data based on ORB-SLAM2 are fused by the EKF method at the situation inference level of the JDL model. The accuracy of position and pose estimation is improved by the complementary advantages of the two methods. (3) the online and offline multi-frame map fusion optimization based on ICP (iterative closed Point) method is implemented. In online mode, the local maps of multiple point clouds are fused together using ICP method to obtain the final laser map. At the same time, the accumulative error of real-time attitude estimation is eliminated as much as possible. In offline mode, based on the online mode, a single local point cloud internal optimization based on ICP method is added.
【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
本文编号：2127807
[Abstract]:Nowadays, many robots need to move in the work environment to complete the task, and map is the basis and key for the mobile robot to understand the surrounding environment and take action. In this paper, a 3D map construction method based on 2D laser is described, which enables the robot system to collect laser data and construct 3D point cloud map in the course of dynamic motion. In order to construct 3D point cloud map by two-dimensional rotating laser sensor, this paper studies the system architecture, sensor data processing, position and pose estimation and so on. The main research contents and results are as follows: (1) the hardware and software architecture of the environment sensing unit is designed, and the multi-sensor time synchronization based on spline smoothing is proposed to solve the problem of multi-sensor original data timestamp inconsistency. Different time synchronization methods are used for different sensors. Generally, linear interpolation is used directly for sensor data, while for visual odometer data with time uncertainty, reprojection error and key frame information are used. The spline smoothing method is used to realize time synchronization. (2) A multi-sensor fusion position and attitude estimation method based on JDL (Joint Directors of Laboratories Model) and extended Kalman filter (EKF) is proposed. In this method, the JDL model is used as the multi-sensor fusion framework, and the binocular visual mileage information and the inertial sensor unit data based on ORB-SLAM2 are fused by the EKF method at the situation inference level of the JDL model. The accuracy of position and pose estimation is improved by the complementary advantages of the two methods. (3) the online and offline multi-frame map fusion optimization based on ICP (iterative closed Point) method is implemented. In online mode, the local maps of multiple point clouds are fused together using ICP method to obtain the final laser map. At the same time, the accumulative error of real-time attitude estimation is eliminated as much as possible. In offline mode, based on the online mode, a single local point cloud internal optimization based on ICP method is added.
【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】
相关期刊论文 前1条
1 苏丽颖,谭民;移动机器人构建地图的研究与发展[J];中国科学院研究生院学报;2002年02期
，本文编号：2127807
本文链接：https://www.wllwen.com/shoufeilunwen/xixikjs/2127807.html
