基于贝叶斯滤波器的移动机器人同时定位与地图创建算法研究

发布时间：2017-12-31 05:15

本文关键词：基于贝叶斯滤波器的移动机器人同时定位与地图创建算法研究　出处：《浙江大学》2016年博士论文　论文类型：学位论文

【摘要】：随着人工智能技术的不断发展,移动机器人在制造工业、国防军事、航天航空、卫生医疗、家庭服务等领域中得到了广泛的应用。移动机器人的同时定位与地图创建是目前机器人学的热点研究课题之一,它是移动机器人实现真正意义上的自主化和智能化的关键前提。由于移动机器人的实际作业环境中存在各种不确定干扰因素,其测量系统的噪声往往具有非高斯重尾分布或者参数先验信息未知等特性。在这些复杂未知环境下,传统的基于贝叶斯滤波估计技术的同时定位与地图创建算法性能受到了严重影响,其定位精度、地图准确性和计算效率无法满足实际应用的需求。为了提高传统算法在复杂未知环境下的估计性能,本文分别对基于高斯滤波器、粒子滤波器和概率假设密度滤波器的同时定位与地图创建算法进行了改进研究,具体内容包括:(1)从贝叶斯滤波估计角度出发,分别对基于高斯滤波器、粒子滤波器和概率假设密度滤波器的同时定位与地图创建算法进行了分析总结,为后续改进相应算法的研究工作提供了理论基础。(2)对测量系统噪声为非高斯重尾分布时的同时定位与地图创建算法进行了研究,提出了一种基于统计线性回归鲁棒优化的同时定位与地图创建算法。首先,基于平方根容积卡尔曼滤波器对联合状态向量进行预测估计。然后,将状态的测量更新方程转换成等价的统计线性回归形式,并利用广义最大似然估计法计算当前测量残差向量对应的增益权值矩阵。最后,通过迭代重加权最小平方法对联合状态的后验均值和协方差平方根因子进行估计。(3)对FastSLAM算法中采样粒子质量差和计算效率低的问题进行了研究,提出了一种改进粒子提议分布估计及粒子重采样方法的UFastSLAM算法。首先,将控制噪声和测量噪声同时增广至机器人位姿状态向量中,并采用平方根转换无味卡尔曼滤波器对最优粒子提议分布进行估计。然后,根据该粒子提议分布对机器人位姿状态进行采样和特征地图的状态更新。最后,在粒子重采样阶段,采用基于KL散度的自适应粒子重采样方法确定当前时刻所需的粒子个数。(4)对环境中同时存在杂波干扰和未知测量噪声方差的同时定位与地图创建算法进行了研究,提出了一种同时估计未知测量噪声方差的随机有限集的同时定位与地图创建算法。首先,将地图特征和测量噪声方差的联合后验强度表示成逆伽马分布和高斯分布的乘积加权和形式,并通过变分贝叶斯近似方法对相关参数进行迭代更新。然后,对联合后验分布混合项进行删减、合并操作以实现地图特征位置状态和个数的估计。最后,根据单特征策略计算粒子权值并对机器人的位姿状态进行估计。
[Abstract]:With the continuous development of artificial intelligence technology, mobile robots in the manufacturing industry, national defense and military, aerospace, health care. Home service and other fields have been widely used. Simultaneous location and map creation of mobile robot is one of the hot research topics in robotics. It is the key premise for mobile robot to realize independence and intelligence in real sense, because of various uncertain interference factors in the actual working environment of mobile robot. The noise of the measurement system often has the characteristics of non-#china_person0# heavy-tail distribution or the unknown prior information of parameters, etc., in these complex unknown environments. The performance of traditional simultaneous location and map creation algorithms based on Bayesian filtering estimation is seriously affected, and its location accuracy is greatly affected. Map accuracy and computational efficiency can not meet the needs of practical applications. In order to improve the performance of traditional algorithms in complex unknown environment, this paper respectively based on Gao Si filter. Particle filter and probability hypothetical density filter are used to improve the algorithm of location and map creation. The specific contents include: 1) from Bayesian filter estimation point of view, respectively based on Gao Si filter. The algorithms of particle filter and probability hypothetical density filter are analyzed and summarized. This paper provides a theoretical basis for the further research of improving the corresponding algorithm. (2) the algorithm of simultaneous location and map creation when the measurement system noise is non-#china_person0# heavy-tailed distribution is studied. A new algorithm of simultaneous location and map creation based on statistical linear regression robust optimization is proposed. Firstly, the joint state vector is estimated based on square root volume Kalman filter. The state updating equation is transformed into an equivalent statistical linear regression form, and the gain weight matrix corresponding to the current measurement residual vector is calculated by using the generalized maximum likelihood estimation method. Finally. The posterior mean and covariance square root factor of joint state are estimated by iterative reweighted minimum square method. The problems of poor sampling particle mass and low computational efficiency in FastSLAM algorithm are studied. In this paper, an improved UFastSLAM algorithm for particle proposal distribution estimation and particle resampling is proposed. Firstly, the control noise and the measurement noise are extended to the pose state vector of the robot at the same time. And the square root transform tasteless Kalman filter is used to estimate the optimal particle proposed distribution. Then, the robot pose state is sampled and the status of the feature map is updated according to the particle proposed distribution. At the particle resampling stage. An adaptive particle resampling method based on KL divergence is used to determine the number of particles needed at the current time. In this paper, the simultaneous location and map creation algorithms of clutter interference and unknown measurement noise variance in the environment are studied. A simultaneous localization and map creation algorithm is proposed for simultaneous estimation of random finite sets of unknown measurement noise variances. The joint posterior strength of map feature and measurement noise variance is expressed as the product weighted sum of inverse gamma distribution and Gao Si distribution, and the relevant parameters are updated iteratively by variational Bayesian approximation. The joint posterior distribution mixed items are deleted and combined to estimate the position and number of map features. Finally the particle weight is calculated according to the single feature strategy and the position and pose state of the robot is estimated.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP242

【相似文献】