【作者】 孙健；

【导师】 余洪山；

【作者基本信息】 湖南大学 ， 控制科学与工程， 2018， 硕士

【摘要】 高精度三维彩色点云地图可重现三维环境场景,是机器人室内定位导航,虚拟现实、增强现实的基础,同时也是人机交互的重要需求之一。将低成本RGB-D传感器用于室内移动机器人环境感知已成为国内外研究热点之一。只有机器人具备了构建三维场景地图的能力,才可能更加自主的在环境中与人进行交互以及执行复杂任务,所以移动机器人三维地图构建是服务机器人迈向智能化的必要条件。点云配准和闭环检测是三维地图构建和机器人运动估计的关键步骤。RGB-D相机可获取稠密的深度信息和对应图像信息,但深度信息精度差、噪声多。因此通常提取匹配图像特征对应的三维点云以降低配准误差和计算量,但受制于场景图像特征的稀疏性。相对于基于空间点位置点云配准方法,三维正态分布变换(3D-NDT)算法将点云分成固定大小的网格体素,基于体素中点云的概率密度实现快速准确匹配,但依赖于初值估计。随着点云地图规模的增大,点云配准不可避免的会产生累积误差,闭环检测对于减少机器人位姿累积误差,避免引入地图冗余变量建立全局一致地图至关重要,基于视觉词袋模型的闭环检测方法仅利用了环境图像特征,对重复结构的场景易引发“感知混淆”,继而导致闭环检测的误检率增加。三维点云地图缺乏对场景描述的高层信息,难以应用于机器人导航。针对以上存在的问题,本文提出了一种面向RGB-D相机的改进3D-NDT点云配准算法,通过自适应设置角点阈值提取特征,以保证场景图像特征稀疏时仍能提取足够特征点;引入运动平滑性约束和双向一致性检验实现误匹配特征点对快速准确剔除;基于匹配特征点云采用PnP算法计算初始配准估计;由于点云位置和数量在空间随机分布,算法引入八叉树模型根据点云密度自适应划分网格体素,采用3D-NDT算法实现特征点云的精确匹配,有效提高了点云表面表达精度以及搜索效率。为克服仅依赖视觉特征的局限性本文建立了一种融合场景视觉特征和几何特征的词袋模型,提高不同环境下词袋模型的鲁棒性。为使机器人能够区分障碍物和可通行区域,对三维点云地图的地面点云进行提取,非地面点云进行聚类,然后将分割好的三维点云转化成八叉树数据结构的三维栅格地图,便于进行机器人进行避障和路径规划。实验表明,改进的点云配准算法对环境的适应性、匹配精度和计算时间方面优于现有三维点云配准方案。与现有的基于视觉词袋模型的闭环检测方法进行对比,本文算法在较高召回率的情况下仍能保证较好的准确率。更多还原

【Abstract】 High precision colored 3D point cloud map can describe the indoor environment in which the robot operates,which is the found ation of localization,navigation,virtual reality and augmented reality,and is also one of the most important requirements for human-computer interaction.Indoor mobile robot environment perception using Low-cost RGB-D sensors has become one of the research hotspots at home and abroad.Only if the robot has the ability to build a 3D scene map,it is possible to interact with people and perform complex tasks in the environment more independently.Therefore,3D mapping is the necessary condition for the service robot to realize intelligence.Point cloud registration and loop detection are key steps in building 3D model of the environment and estimating the robot pose.RGB-D cameras capture RGB images along with per-pixel depth information,but the depth estimates are low accuracy and noisy.Therefore,it is the immediate local neighborhoods of the detected local visual feature points that reduce the registration error and computation,but which is limited to the rich visual features.Relative to the point to point registration,the Three-Dimensional Normal Distributions Transform(3D-NDT)algorithm divide the point clouds into fixed 3D voxel grid cell with a fixed size,each 3D voxel grid cell is described by PDF to achieve rapid and accurate registration,although it depends on the good initial estimate.With the scale of point cloud map increasing,point cloud registration will inevitably produce cumulative error.Loop-closing detection is very important for reducing the accumulative error of robot trajectory and avoiding the introduction of redundant structures to ensure global consistency of maps.The methods based on the visual bag-of-words detect loop-closing can easily lead to perceptual confusion by extracting the visual features,which lead to the increase of false detection rate in loop-closing.3D point cloud map lacks high-level information to describe scenes,so it is difficult to use in robot navigation.Based on the above problems,this thesis presents a RGB-D registration method based on combinations between local visual feature and improved 3D-NDT,Firstly,FAST feature detection based on self-adaptive threshold selection is proposed to ensure enough feature points in weakly textured environment.Secondly,the fast and robust false matching points elimination algorithm is proposed by incorporating the smoothness constraints and bidirectional consistency matching.Based on the matching feature point 3D position,the initial estimate is obtained by PnP.While the distribution and number of point cloud is random in the space,the voxels based on octrees are used to subdivide point cloud by the distribution-density,then the 3D-NDT guarantee an accuracy registration.It can accurately describe the surface details of the point cloud and improve data search effectivity.In order to overcome the limitation that only depends on the visual features,a bag-of-words model which combines the visual features and geometric features of the scene is built to improve the robustness of the loop-closing algorithm in different environments.In order to make the robot understand the obstacle and the accessible areas,this thesis segments the point cloud map into two parts,and then transform them into the octree-based mapping framework OctoMap,which is essential for robot to carry out the obstacle avoidance and path planning.Experiments show that the improved point cloud registration algorithm outperforms other state of art of 3D point cloud registration methods on the adaptability to environment,precision and computation cost.Compared with the loop detection methods based on visual bag-of-words model,the algorithm can guarantee better accuracy in the case of higher recall rate.更多还原