【作者】 刘丹；

【导师】 段建民；

【作者基本信息】 北京工业大学 ， 控制科学与工程， 2018， 博士

【摘要】 智能车辆(Intelligent Vehicle,IV)是智能交通系统的重要组成部分。为了能够彻底将人类从繁重的驾驶过程中解放出来,有效避免交通事故的发生,智能车辆需要完全实现无人驾驶。同时定位与建图(simultaneous localization and mapping,SLAM)技术作为智能车辆实现自主导航的关键技术,能够为车辆在定位系统失效的未知环境中实现无人驾驶提供必要的条件。为此,论文分别从基于概率的SLAM方法、SLAM中的数据关联方法和基于扫描匹配的地图构建与定位方法三个方面出发,对智能车同时定位与建图的关键技术进行了深入的研究,主要内容如下:(1)针对智能车辆SLAM系统,定义了其研究中所需要的坐标系,建立了车辆运动学模型、传感器观测模型、环境地图模型以及数据关联模型,并基于上述模型给出了智能车SLAM问题的概率模型,为智能车同时定位与建图关键技术的研究搭建了统一平台。(2)针对几何特征地图中自然实体路标的提取问题,研究了一种基于激光雷达数据的圆型特征提取方法。该方法可根据雷达数据点的距离信息和角度信息提取环境中实体路标的中心和直径。通过Victoria Park数据集验证了圆型特征提取方法的有效性,从而为后续SLAM算法中的几何地图构建提供特征提取方法。针对基于扩展卡尔曼滤波的SLAM(EKF-SLAM)算法易受非线性模型的不确定性和误差统计的不确定性影响,基于强跟踪滤波思想,提出了一种自适应渐消EKF-SLAM算法。该算法为后续章节核心算法的研究奠定了理论基础。(3)为了解决基于粒子滤波器的快速同时定位与建图算法(Fast simultaneous localization and mapping,FastSLAM)一致性差、估计精度随粒子的退化和贫化逐渐降低的问题,提出了一种基于改进粒子建议分布函数和部分重采样策略的FastSLAM算法。在该算法中,设计了强跟踪平方根中心差分卡尔曼滤波(strong tracking square root central difference Kalman filter,STSRCDKF);在车辆位姿估计阶段利用STSRCDKF获取可自适应调节的建议分布函数,使其更贴近粒子后验概率分布,进而提高粒子采样精度;在地图估计阶段,采用STSRCDKF实现对环境路标位置的估计,提高建图精度;在重采样阶段,采用部分重采样策略降低粒子集的退化和贫化现象,提高算法的一致性。实验结果证明了提出算法在鲁棒性、一致性和估计精度方面的优势。(4)在SLAM中,数据关联作为状态估计的前提和基础,是保证定位与建图过程收敛的核心和关键。针对目前SLAM中所应用的数据关联算法无法同时保证计算复杂度低和关联正确率高的问题,提出了两种不同的联合数据关联算法。首先,在联合兼容分枝定界(Joint Compatibility Branch and Bound,JCBB)算法的基础上,提出了一种基于聚类分组策略和中心差分联合兼容准则的数据关联算法,该算法有效解决了JCBB算法易受线性化误差影响和复杂度高的问题,能够在获得准确关联结果的同时,降低SLAM算法的复杂度。其次,根据联合最大似然准则将SLAM数据关联问题转化为一种组合优化问题,采用一种基于跳跃行为和自适应步长改进的人工鱼群算法搜索最优数据关联解。实验结果表明提出的两种关联算法能够为提高智能车SLAM的实时性和准确性提供可靠的保障。(5)为了实现车辆准确自定位且构建描述环境细节的稠密特征地图,提出了一种基于扫描匹配和粒子滤波器的点云地图创建与定位方法。采用强跟踪平方根中心差分粒子滤波融合基于迭代最近点(Iterative Closest Points,ICP)匹配的定位结果和基于里程计的定位结果,有效避免了扫描匹配定位过程中的累积误差对车辆位姿估计和地图更新的影响。在点云地图构建阶段,通过寻找当前扫描数据点与参考数据点之间的对应关系,将存在对应关系的点根据各自的权重值进行融合;最后基于车辆的全局位姿完成点云地图的拼接。实验结果表明提出的方法不仅能够实现车辆的准确定位,而且构建的点云地图可以为智能车自主驾驶提供细致的环境信息。更多还原

【Abstract】 Intelligent vehicle(IV)is a major part of intelligent transportation system.In order to completely emancipate human beings from the heavy driving process and effectively avoid traffic accidents,intelligent vehicle needs to be achieved self-driving.The simultaneous localization and mapping(SLAM)is a key technology of autonomous navigation for intelligent vehicle,it can provide the necessary conditions for the vehicle to realize self-driving in an unknown environment where the positioning system fails.Hence,from three aspects,the SLAM method based on probability,the data association method of SLAM and the SLAM method based on scan matching,respectively,the key technologies of SLAM for intelligent vehicle are studied.The major contents are summarized as follows:(1)Due to the SLAM system of intelligent vehicle,the coordinates needed in the research of SLAM is defined.The vehicle kinematic model,sensor observation model,environmental map model and data association model are developed.The probability model of the SLAM problem for intelligent vehicle is given based on the above models,which have built a unified platform for the research on key technologies of SLAM.(2)Aiming at the problem of extracting the natural entity landmark in the geometric feature map,a circular feature extraction method based on laser radar data is studied.The center and diameter of the entity landmarks are extracted based on the distance information and angle information from laser radar data in the proposed extraction method.The effectiveness of circular feature extraction is verified by Victoria Park dataset,which provides a feature extraction method for building geometric map in subsequent SLAM algorithm.Due to the SLAM algorithm based on the extended Kalman filter is susceptible to the uncertainty of the nonlinear model and the uncertainty of the error statistics.Based on the idea of strong tracking filter,an adaptive fading EKF-SLAM algorithm is proposed.This proposed algorithm lays a theoretical foundation for the research of the core algorithms in subsequent chapters.(3)In order to solve the problem of poor consistency and the estimation accuracy decreasing with the degradation and impoverishment of particles in FastSLAM,a FastSLAM algorithm based on the improved proposal distribution and partial resampling strategy is proposed.In the proposed algorithm,the strong tracking square root central difference Kalman filter is designed.STSRCDKF is used to design an adaptive adjusting proposal distribution of the particle filter in the estimation stage of vehicle’s pose.Because the proposal distribution is very close to the posterior probability distribution of particles,hence the sampling precision of particle is improved.In the stage of map estimation,STSRCDKF is used to estimate the environmental landmark to improve the accuracy of building map.In the resampling stage,a partial resampling strategy is adopted to reduce the degradation and impoverishment of the particle set,and the consistency of the algorithm is improved.The experimental results demonstrate the advantages of the proposed algorithm in robustness,consistency and estimation accuracy.(4)In SLAM,data association is the precondition and basis of state estimation,which is the core and key to ensure the convergence of localization and mapping.In order to solve the problem that the association algorithms used in SLAM can not simultaneously guarantee low computational complexity and high association correct rate.Two different joint data association methods are proposed.Firstly,a data association method based on clustering strategy and central difference joint compatibility criterion is proposed based on joint compatibility branch and bound(JCBB)algorithm.This method effectively solves the problem that JCBB algorithm is susceptible to linearization error and high complexity.It can reduce the complexity of the SLAM algorithm while obtaining the accurate association results.Secondly,according to the joint maximum likelihood criterion,the SLAM data association problem is transformed into a combinatorial optimization problem.An artificial fish swarm algorithm based on jump behavior and adaptive step-size is applied to search for the optimal data association solution.The experimental results show that the proposed association methods can provide a reliable guarantee for improving the real-time and accuracy of SLAM for intelligent vehicle.(5)In order to realize vehicle’s self-localization and build a dense feature map for describing the details of the environment,a SLAM method based on scan matching and particle filter is proposed.Strong tracking square root central difference particle filter is used to fuse location results based on ICP matching and location results based on odometer.The effect of cumulative error on vehicle’s pose estimation and map updating is effectively avoided in the process of scan matching.In the construction stage of point-cloud map,the corresponding relationship between the current scan points and the reference points is found,and the points that have corresponding relationships are fused according to their respective weight values.Finally,the point-cloud map stitching is completed based on the global pose of the vehicle.Experimental results show that the proposed method can accurately estimate vehicle’s pose,and the point-cloud map can provide detailed environmental information for autonomous driving of intelligent vehicle.更多还原