节点文献
室内未知环境线段特征地图构建
Line Feature Map Building for Unknown Indoor Environments
【作者】 熊蓉;
【导师】 褚健;
【作者基本信息】 浙江大学 , 控制科学与工程, 2009, 博士
【摘要】 未知环境地图构建是机器人和人工智能研究领域的一个重要课题,是移动机器人在未知环境中自主完成侦察、勘探、搜索、救援、导航等各项工作的基础,对提高机器人的智能性、促进机器人进入人类日常生活、为人类服务具有重要的研究意义和应用价值。未知环境地图构建问题面临着同时定位与地图构建(SLAM,SimultaneousLocalization and Mapping)方法、高维数据、数据对应、环境的动态特性和环形特性、以及自主探测规划等多方面的挑战。结合室内环境结构性特点,本文采用描述障碍物轮廓的线段为特征,通过对信息处理、同时定位与地图构建和探测规划问题的深入研究,建立了一套相对完整的线段特征地图构建方法。同时,为了提高地图构建机器人快速运动能力和在窄小空间中的灵活运动能力,本文研究了四轮全方位移动机器人的运动建模与控制问题,为后续研发奠定了基础。本文主要研究工作包括:(1)提出了结合哈夫变换、同线性判断和最小二乘法的混合线段拟合方法。该方法可有效提高拟合精度,实现对环境信息精确而简洁的描述。(2)提出了寻找测量数据与线段特征最佳相合的增量式同时定位与地图构建方法。该方法将SLAM分解为局部地图构建、机器人位姿估计和地图合并三个循环步骤。在地图估计步骤,利用最小二乘法迭代寻找当前测量数据与已构建地图中线段特征的最佳相合实现机器人位姿估计,并通过去除不当匹配和引入加权矩阵来减小测量误差、特征拟合误差和已构建地图中的不确定性对位姿估计的影响。在地图合并步骤,根据所估计位姿,合并当前观测得到的局部线段特征地图和已构建全局线段特征地图实现地图的更新。该方法避免了高斯噪声假设,降低了对数据关联错误的敏感性,具有较小的匹配运算量,可以在线实时构建线段特征地图。在较复杂的实验室环境和较大的楼道环境中的实验证明了算法的有效性和鲁棒性。(3)提出了基于点线相合和粒子滤波的FastLineSLAM方法。该方法采用粒子实现机器人路径的多假设,在粒子中采用基于点线相合的增量式SLAM方法进行地图的更新估计。在粒子采样过程中,利用点线相合的位姿估计方法缩小采样空间;通过基于相合关系的粒子权重更新方法降低计算复杂度;通过选择性重采样抑制粒子滤波常见的退化现象和采样枯竭问题。实验结果证明该算法解决了点线相合SLAM方法存在残差累积、后期难以校正的问题,可良好解决环形环境和动态环境下的室内地图构建,克服了传统粒子滤波SLAM方法存在存储空间负荷高、计算量大的缺陷,所需粒子数和存储空间均较少。(4)提出了一种基于线段特征方向引导的探测规划算法。在利用线段特征生成候选视点的同时,赋予候选视点继承线段特征的方向属性;通过定义具有起始位置和探测方向属性的探测区域,将探测环境表示为具有递进关系的探测区域树;根据候选视点与探测区域的隶属关系,通过探测区域的方向引导或者候选视点的观测方向引导启发式地搜索下一步最佳探测位姿。实验表明,所提方法可有效确保候选视点的可定位性,提高搜索效率,减少来回往复的运动现象。(5)研究了实时动态环境中四轮全方位移动机器人的运动控制和轨迹规划。通过对机器人运动学和动力学特性的分析,给出了四轮全方位移动机器人的控制模型,并根据方程特性对其进行了合理的简化,使得计算量有效减少。同时采用Bang-Bang控制规划出时间最优的机器人运动轨迹。通过轨迹规划和控制模型的结合达到了实时控制的效果。实验证明了模型的正确性和算法的有效性。
【Abstract】 Robotic mapping is an essential issue in robotics and artificial intelligent.It acts as the base of reconnaissance,search,rescue,navigation et al for mobile robot working in unknown environment.Thus its research and application is much significant to improve the intelligence of robot and promote the process that robots service for human in daily life.The main challenges of robotic mapping exist on SLAM(Simultaneous Localization and Mapping),high dimensionality,data association,dynamic and loop features in environment and autonomous exploration.This dissertation addresses to robotic mapping in unknown indoor environments.Considering the structure characteristic of indoor environment,we adopt line that describes the profile of obstacle as feature.With the research on information process,SLAM and exploration,an integrated map building approach is established.In addition, to boost the mobility of the robot for map bilding in the narrow environment,the motion modeling and optimal control of omni-directional mobile robots isstudied, which provides a base for future work.The whole paper includes the following detail:(1) A combining method of Hough transform,coincided-line detecting and Least Square is presented and used to fit line segment from measurement.This method boosts the fitting precision,thus the environment can be described accurately and succinctly.(2) We proposed an incremental SLAM approach based on the best congruence between dot data in the current measurement and line segments in the previously-built map.Each iteration of SLAM consists of 3 stages:local map building,robot pose estimating,and map integrating.In pose estimating,least square method is used iteratively to obtain the best correspondence between the measurement and the half-baked map.Removing improper match and defining weighted matrix are both implemented to reduce the errors of measurement,line fitting and previously-built map.In map integrating,with the pose estimated,the map is updated by fusing the local map built from current measurement and the previously-built map.This method avoids the hypothesis of Gauss noise as EFK-SLAM and reduces the sensitivity to the error of data corresponding.It can work online with a low computation load in match.Experimental results with real data demonstrate the approach is effective and robust for indoor environment mapping.(3) We further proposed an approach called FastLineSLAM by introducing the incremental SLAM algorithm based on dot-line congruence into particle filter. In the approach,each particle carries an assumption on robot path and employs the SLAM algorithm based on dot-line congruence to update the map. Both the motion and the observation information are considered in the importance function by using the dot-line congruence method to estimate the pose of robot.The weight of the particle is updated according to the congruence between current measurement and segment features in previously-built map.The wrong particles resulted from mis-matching or error accumulation are filtered with selective resampling.Experimental results with real data demonstrate the approach is effective and robust for mapping dynamic and loop indoor environment by solving the residual error existing in the incremental SLAM algorithm based on dot-line congruence.Both of the particle number and memory are quite lower than the existing mapping methods using particle filter.(4) We presented a line-feature-guided exploration approach to find the next best pose of view.The candidates of view are not only generated from line features but also inherited the orientation property of line features,which is used to guide the exploration.The environment is divided into a sequence of exploration space which is defined as a subspace with an exploration direction and a start position.According to affiliation between candidates and exploration spaces,an heuristic NBV search is implemented by obeying the direction guide of both the candidate and the exploration space.Experimental results demonstrate the proposed approach ensures the localization of candidates of view and is efficient for active map building in indoor environment to get a good complete coverage with few criss-crossing motion.(5) We studied the motion control and path planning of the omni-directional mobile robot.Through analysis the characteristic of the kinematics and dynamics of the omni-directional robot which equipped four omni-directional wheels,its motion control model is provided.Then the model is simplified rationally according to the feature of model equations,which reduces the cost of computation effectively.Using Bang-Bang control,time-optimal trajectory generation method also is carried out and produced a real-time effect integrated with the motion control model.The effectiveness of the method has been demonstrated by experiments.
【Key words】 Unknown Indoor Environment; Map Building; SLAM; Exploration; Motion Control;