不依赖GPS定位理论及方法研究

发布时间：2018-10-04 20:52

【摘要】：智能车,又称无人驾驶智能车,是近些年汽车领域和车辆领域科学研究和产品发展的主要方向。智能车为了能够完成人们期望的各类任务,定位精度和定位的可靠性是其最基本的问题,也是相对复杂的问题。本文主要研究GPS失效环境下的新型定位算法,旨在补偿智能车在缺少GPS信号情况下也可以获得精确的定位信息。首先考虑在实际应用过程中,多数情况下并不能对车辆进行有效的建模或者车辆上并不方便安装里程计,此时传统基于里程计的车辆运动学模型将无法应用于车辆定位中。为了解决缺少里程计情况下的车辆定位问题,本文提出一种车辆运动状态估计模型。提出算法通过该模型估计车辆位置、姿态及其运动状态(如速度)。提出算法用估计的运动状态代替里程计,进而估计车辆的位置。为了验证算法的效果,本文将提出算法同基于模型车辆定位算法进行比较,先后通过仿真数据和悉尼大学维多利亚数据库数据进行对比。实验结果表明,提出算法可以达到基于里程计的算法同等的定位精度。之后本文研究了视觉里程计并提出一种新的视觉里程计算法。该算法通过解耦估计旋转-平移增强该视觉里程计在动态环境下的性能。理想的视觉里程计通过观测静态环境特征实现视觉系统的运动估计,但是实际环境中不可避免的存在动态特征。因而如何消除动态特征,减少其对视觉里程计性能的影响,是提高里程计性能的有效途径。本文通过立体视觉系统将特征点划分成“远点”和“近点”分别处理：在一致随机估计(RANSAC)框架下,采用“远点”估计视觉系统的姿态；进而在姿态已知的前提下,使用“近点”计算摄像机平移。这样通过姿态约束降低了近距离运动物体对视觉里程计的影响。实验表明,在实际道路环境中,本文基于旋转-平移解耦估计的算法较之传统同时估计旋转平移的算法,能有效剔除动态特征,提高视觉里程计的性能。为了获得更高精度的定位结果,本文研究了车辆同时定位与地图构建(Simultaneous Localization And Mapping,SLAM)问题。SLAM是将车辆和环境地图看成一个整体,通过估计车辆位置构建增量式地图,再根据地图估算车辆位置。尽管已有的SLAM算法能够有效降低累积误差,但是这些算法大都存在线性化误差过大的问题。本文提出两种新的SLAM算法：平方根正交容积变换卡尔曼滤波器的SLAM算法(SCQKF-SLAM)和基于正交容积粒子滤波器的SLAM算法(CQFastSLAM)。这两个算法均通过融合正交容积变换(cubature rule)和高斯-拉格朗日规则(Gauss-Laguerre quadrature rule)求解出更高的非线性精度。通过综合考虑计算代价和计算,SCQKF-SLAM和CQFastSLAM都是有效降低SLAM线性化精度。后续仿真和数据库实验表明,在大规模环境地图中,提出算法都有效提高了车辆定位精度。
[Abstract]:Smart car, also called driverless smart vehicle, is the main direction of scientific research and product development in automobile field and vehicle field in recent years. In order to accomplish all kinds of tasks, positioning accuracy and reliability are the most basic and relatively complex problems. In this paper, a new localization algorithm in GPS failure environment is studied in order to compensate the intelligent vehicle for obtaining accurate location information without GPS signal. First of all, in the practical application process, most of the vehicle can not be effectively modeled or installed on the vehicle is not convenient to install mileage meters, at this time, the traditional vehicle kinematics model based on mileage meter will not be applied to vehicle positioning. In order to solve the problem of vehicle location in the absence of mileage, a vehicle motion state estimation model is proposed in this paper. An algorithm is proposed to estimate the vehicle position, attitude and motion state (such as velocity) by the model. An algorithm is proposed to estimate the vehicle position by replacing the mileage meter with the estimated motion state. In order to verify the effectiveness of the algorithm, the proposed algorithm is compared with the model-based vehicle location algorithm, and the simulation data are compared with the Victorian database data of the University of Sydney. Experimental results show that the proposed algorithm can achieve the same positioning accuracy based on odometer. Then, the visual mileage meter is studied and a new method for calculating visual mileage is proposed. The algorithm improves the performance of the visual odometer in dynamic environment by decoupling estimation of rotation and translation. The ideal visual mileometer can estimate the motion of the visual system by observing the static environmental features, but the dynamic features are inevitable in the actual environment. Therefore, how to eliminate the dynamic features and reduce the impact on the performance of the visual odometer is an effective way to improve the performance of the odometer. In this paper, the feature points are divided into "distant points" and "near points" by stereo vision system. In the framework of uniform random estimation of (RANSAC), the "distant points" are used to estimate the attitude of the visual system, and then, when the attitude is known, Use close Point to calculate camera translation. In this way, the influence of close moving object on visual odometer is reduced by attitude constraint. The experimental results show that the proposed algorithm based on rotation-translation decoupling estimation can effectively eliminate the dynamic features and improve the performance of visual odometer compared with the traditional simultaneous estimation of rotation-translation algorithm in the actual road environment. In order to obtain more accurate positioning results, the problem of simultaneous vehicle location and map construction (Simultaneous Localization And Mapping,SLAM) is studied in this paper. SLAM regards vehicle and environment map as a whole, and builds incremental map by estimating vehicle location. Then estimate the location of the vehicle according to the map. Although the existing SLAM algorithms can effectively reduce the cumulative error, most of these algorithms have the problem of excessive linearization error. In this paper, we propose two new SLAM algorithms: SLAM algorithm for square root orthogonal volume transform Kalman filter (SCQKF-SLAM) and SLAM algorithm (CQFastSLAM). Based on orthogonal volume particle filter. Both of these algorithms can solve higher nonlinear accuracy by merging orthogonal volume transform (cubature rule) and Gauss Lagrangian rule (Gauss-Laguerre quadrature rule). Both SCQKF-slam and CQFastSLAM are effective in reducing the linearization accuracy of SLAM by considering the computational cost. The following simulation and database experiments show that the proposed algorithm can effectively improve the accuracy of vehicle location in large-scale environmental maps.
【学位授予单位】：北京交通大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：U495;U463.6

【相似文献】