【作者】 李亮；

【导师】 杨明；

【作者基本信息】 上海交通大学 ， 控制理论与控制工程， 2018， 博士

【摘要】 智能车是结合人工智能、机器人学、控制规划等多门学科的综合应用研究。涉及到目标检测、场景理解、高精度定位、路径规划、运动控制等众多关键技术。在支撑智能车的众多技术中,高精度定位是其中最为基础且关键的。智能车所行驶的环境大多为已知环境,因此基于点云匹配的定位成为目前主流的定位方法。基于点云匹配的定位中最常用的两种传感器包括视觉传感器与激光雷达,由于视觉传感器对于光线变化、天气变化等状况的鲁棒性较差,基于激光雷达的定位方案越来越受到青睐。但基于激光雷达的定位方法由于地图存储量大、环境表示效率低、地图管理难度大、地图匹配算法鲁棒性差等缺点使得其应用受到制约。为解决以上问题,本文提出了一种基于点云匹配的定位框架。为解决点云匹配中的精度与鲁棒性问题,本文提出了鲁棒的点集配准算法。具体工作如下:针对点云地图数据存储量大,不适用于大范围定位的问题。本文对拼接点云进行三维栅格化与平面压缩处理生成感知地图。首先将点云转换成三维占据栅格图,并采用八叉树的格式进行存储。这不仅可以降低存储量、抽取有用信息,还可以通过多分辨率格式,提高定位的鲁棒性与实时性。在道路两侧结构特征明显的区域,进一步压缩八叉树地图,得到街景压缩地图。街景压缩地图的存储量可降低到原始点云存储量的千分之一。本文给出了基于点云匹配的智能车定位算法框架。并且将地图匹配作为观测,融合车辆运动数据,增强定位的鲁棒性。针对对应点选取对初始值依赖较大的问题,本文提出基于图同构的对应点选取算法。首先根据点与点之间的几何关系构建图,并得到相应的邻接矩阵。由此构建的两个点集的邻接矩阵为同构矩阵,通过寻找一个置换矩阵建立二者的同构关系,进而得到对应点关系。为此本文构建了一个优化问题,并通过蒙特卡洛方法求解。采用交替方向乘子算法在蒙特卡洛优化的观测更新中对粒子权重进行更新。实验结果表明当初始值、噪声等干扰逐步增大时,该算法的对应点选择正确率相比于传统算法可提高50%以上。针对点云匹配容易受噪声、外点等干扰问题,本文提出了基于容积卡尔曼滤波的点集配准算法。首先,本文将点集配准问题用状态空间模型表示,由此将点集配准问题转换为滤波问题。由于该问题是高度非线性的,常规的滤波方法精度难以满足要求。因此本文采用容积卡尔曼滤波方法,利用数值方法近似滤波中非线性部分。相比于扩展卡尔曼滤波和无迹卡尔曼滤波,容积卡尔曼滤波可将非线性近似到三阶。实验结果表明,该方法对于噪声、外点、初始值都具有较强的鲁棒性。相比于当前广泛应用的点集配准算法,当噪声大小达到30%时,该方法匹配成功率可提高20%以上;当外点数量达到50%时,该方法匹配成功率可提高30%以上;当结构性缺失大小达到50%时,该方法匹配成功率可提高30%以上。采用容积卡尔曼滤波求解点集配准问题,提高了算法对于单一误差源的鲁棒性与精度。但卡尔曼滤波没有考虑两个数据之间的关联性,多种相关联的误差源同时存在时,算法精度有所降低。为此本文提出了容积分离协方差交集滤波算法求解点集配准问题。该方法将滤波中的方差分为两部分:独立方差与非独立方差,并分别进行处理。这对于处理点集配准中常见两类误差(独立误差与非独立误差)具有较强的鲁棒性。当同时存在多种误差源时,相比于容积卡尔曼滤波的点集配准方法,容积分离协方差交集方法的鲁棒性表现得更好。实验结果表明,当两种误差源同时存在且误差较大时,基于容积分离协方差交集滤波的点集配准比基于容积卡尔曼滤波的点集配准的匹配误差降低40%以上。针对点集配准高度依赖对应点选择正确率的问题,本文提出了一种直接匹配算法。该算法不建立两个点集中的对应点关系,而是将两个点集都表示成高斯混合模型,通过优化两个高斯混合模型的相似度来达到点集配准的目的。为此本文提出了签名二次距离来衡量高斯混合模型相似度,并将其用于点集配准算法。相比于欧式距离、相对熵、核相关函数等,签名二次距离计算相似度函数在输入的两个高斯核上的期望,因此能更好地衡量高斯混合模型间的相似度。对于签名二次距离的优化,采用最大期望算法进行迭代优化,保证算法的收敛性。实验表明该算法准确度相比于对应点算法有所提升,但算法复杂度更高,因此适用于地图创建等离线应用。更多还原

【Abstract】 Research on intelligent vehicles is a cross domain which involves artificial intelligence,robotics,motion planning,control,etc.Among the techniques that make the vehicle intelligent,high-precision localization is basic and important.Because the environment intelligent vehicles travel in is priorly known in most cases,point cloud matching based localization has become the mainstream manner.Camera and LIDAR are two types of sensor which are used most widely in mapping and localization.As vision-based method is not robust to light condition variance,weather changes,etc,LIDAR based method becomes more and more favored by the researchers.For LIDAR based method,there are also some problems that need to be solved:It needs large storage to store the sensor map;The efficiency for environment representation is low;Map management is difficult;Map matching algorithm is not robust to local minimum,noise and outliers.In terms of these problems,a point cloud matching based localization and several point set registration algorithms for various scenes are proposed in this thesis.Details of these works are as following:In order to decrease storage needed to store the map,the stitched point cloud is post-processed to generate sensor map,precision and information of the map are not affected at the same time.For general environment,the point cloud is rasterized into 3D occupancy grid map which is stored in the form of octree.By doing this,it could not only decrease map storage,extract useful information,but also improve localization robustness via the multi-resolution octomap.In the structural areas,the octomap is compressed further to obtain compressed road scene map.Compared with raw point cloud,storage for compressed road scene map decreases to one thousand of the point cloud.This thesis gives the details of framework for point cloud matching based localization.The vehicle motion data are also fused and the map matching result is used as observation,which could improve robustness of localization.For the problem that the correspondence establishment relies heavily on initialization,a correspondence establishment algorithm between point sets based on graph isomorphism is proposed.The graph is established based on the geometric relations between points,after which the adjacency matrix could be obtained.Adjacency matrix of the two point sets are isomorphic,correspondence between points cloud be obtained through the permutation matrix between the two graphs.In light of the above,an optimization problem is constructed and solved by Monte Carlo method.In measurement update of the Monte Carlo optimization,weights of the particles are updated using Alternating Direction Method of Multipliers.Experimental results demonstrate that robustness to noise,outliers and misalignment initialization of the proposed methods outperforms the traditional methods with improvement of correspondence accuracy larger than50%.For the problem that the registration algorithm is not robust to noise,outliers,etc.,a point set registration algorithm based on Cubature Kalman Filter(CKF) is proposed.The point set registration is modeled using model space,then point cloud matching is converted to a filtering problem.Because the model is highly nonlinear,precision of the traditional methods is not high enough.Herein CKF is adopted to deal with the nonlinear part by numerical calculation.Compared with extended Kalman filter,unscented Kalman filter,CKF cloud approximate the nonlinearity for three orders.Experimental results show that the proposed method is robust to noise,outliers and misalignment initialization.It also outperforms some state-of-the-art methods in terms of accuracy and robustness.When the noise is 30%,the successful matching rate of the proposed method could be improved 20%.When the noise is 50%,the successful matching rate is improved by 30%.When the missing partial structures is 50%,the successful matching rate is improved by 30%.CKF based point set registration does not consider correlation between the error sources if there are more than one error source.Due to this,point set registration based on Cubature Split Covariance Intersection Filter(CSCIF) is proposed in this thesis.The proposed method divides covariance in filtering into two parts: independent and dependent,which are dealt with differently.By doing this,it is robust to two common errors in point set registration(independent error and dependent error).Compared with CKF based point set registration,Robustness to various types of error sources of CSCIF based point set registration has been improved substantially.Compared with state-of-the-art point set registration methods,robustness of the proposed method is improved further.When there are two types of error sources that exist simultaneously,error of CSCIF declines 40% compared with the CKF based point set registration.For the problem that the point set registration relies on the correctness percentage of point correspondence,a direct point set registration algorithm is presented in this thesis.The proposed method does not establish correspondence between point sets,but represents the point set using Gaussian Mixture Model(GMM).Point set registration is achieved by optimizing similarity of the two GMMs.Signature Quadratic Form Distance(SQFD)is used to measure the similarity between the two GMMs.Compared with L2 distance,K-L distance,kernel correlation function,etc.,SQFD computes expectation of the similarity function on the two input Gaussian kernels,which could measure similarity between GMMs better.For optimization of SQFD,Expectation Maximization is adopted in an iterative manner to guarantee convergence of the proposed algorithm.Experimental results demonstrate that precision of the GMM-SQFD based point set registration is improved compared with the correspondence-based point set registration.While the runtime complexity of GMM-SQFD based point set registration is higher than the correspondence-based point set registration.Thus,it is more suitable for the offline tasks,e.g.,map generation.更多还原