【作者】 王立银；

【导师】 赵铁军； 隋春平；

【作者基本信息】 沈阳工业大学 ， 工程硕士（专业学位）， 2023， 硕士

【摘要】 随着工业生产的智能化和生产线的柔性化,致使目前全方位自动导航车的工作环境越来越复杂,非结构化的环境也使得传统的导航方式不再适用。本文通过分析激光与视觉两大类SLAM技术,针对其单一传感器SLAM算法建图精确性与系统鲁棒性不足的问题,通过对激光与视觉结合的SLAM算法前端及后端的改进,使得改进后的算法适用性强、可靠性高、鲁棒性好。在建立好的地图上运用改进RRT算法为全方位车提供了路径规划的方法。具体研究内容如下:对全方位车的运动模型进行了推导,针对实际任务的不同要求对全方位车的车体进行了设计,对全方位车使用的传感器进行选型,对相机和雷达的观测模型进行建立。基于深度相机进行了标定实验,并对相机和激光雷达的相对关系进行推导。对视觉里程计间接法的三种特征点法进行筛选,验证之后选择了ORB特征点法;在匹配方案中通过过滤排序思想对FLANN算法进行改进,使得改进算法提高了匹配的正确率并减少了匹配时间;针对激光里程计的激光点云中存在的噪声问题通过评判函数进行评判后,对缺失点运用加权最小二乘法拟合后进行插值的算法进行改进,保证了点云降噪的有效性。对ORB-SLAM2框架前端和后端进行改进。在前端将相机处理的特征点云进行反投影处理,运用筛选函数对高度不同的点云进行筛选,对激光点云与处理后的相机特征点云分别利用加权最小二乘法和线性法进行拟合,使得改进后算法提升了建图的精确性。在后端将改进的特征点匹配算法融合于框架并运用数据集对混合SLAM算法估计精度进行验证,得出其估计精度的优越性。在全局路径规划上对贪心RRT算法扩展随机点时运用偏移目标的算法进行改进,进行实验验证后得出此算法较原算法时间和路径长度都优于原算法。利用系统搭建了仿真环境对SLAM算法进行了仿真,利用ROS系统以及选型的传感器和底盘系统搭建了实验平台,选取室内场地模拟工况进行实际定位、建图和全局路径规划实验,验证了改进的SLAM算法建图的精确性以及路径规划算法的有效性。更多还原

【Abstract】 With the intelligence of industrial production and the flexibility of production lines,the working environment of omnidirectional automatic navigation vehicles is becoming increasingly complex,and the unstructured environment also makes traditional navigation methods no longer applicable.This article analyzes the two major categories of SLAM technologies,laser and vision,and addresses the issue of insufficient accuracy and system robustness in the construction of single sensor SLAM algorithms.By improving the front-end and back-end of the SLAM algorithm that combines laser and vision,the improved algorithm has strong applicability,high reliability,and good robustness.The use of improved RRT algorithm on established maps provides a path planning method for omnidirectional vehicles.The specific research content is as follows:The motion model of the omnidirectional vehicle was derived,and the vehicle body was designed according to different requirements of actual tasks.The sensors used in the omnidirectional vehicle were selected,and the observation models of cameras and radar were established.A calibration experiment was conducted based on a depth camera,and the relative relationship between the camera and the Li DAR was derived.Three feature point methods for indirect visual odometry were screened and validated,followed by the ORB feature point method;In the matching scheme,the FLANN algorithm is improved by filtering and sorting ideas,which improves the accuracy of matching and reduces matching time;After evaluating the noise problem in the laser point cloud of the laser odometer through the evaluation function,the algorithm for interpolating missing points using weighted least squares fitting was improved to ensure the effectiveness of point cloud noise reduction.Improve the front-end and back-end of the ORB-SLAM2 framework.At the front end,the feature point cloud processed by the camera is backprojected,and a filtering function is used to filter point clouds with different heights.The laser point cloud and the processed camera feature point cloud are fitted using weighted least squares and linear methods,respectively,to improve the accuracy of the improved algorithm.In the backend,the improved feature point matching algorithm is integrated into the framework and the estimation accuracy of the hybrid SLAM algorithm is verified using a dataset,obtaining the superiority of its estimation accuracy.In global path planning,the greedy RRT algorithm was improved by using the offset target algorithm when expanding random points.After experimental verification,it was found that this algorithm outperformed the original algorithm in terms of time and path length.A simulation environment was built using the system to simulate the SLAM algorithm.An experimental platform was built using the ROS system,selected sensors,and chassis systems.Indoor site simulation conditions were selected for actual positioning,mapping,and global path planning experiments,verifying the accuracy of the improved SLAM algorithm in mapping and the effectiveness of the path planning algorithm.更多还原