【作者】 郭剑辉；

【导师】 赵春霞；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2008， 博士

【摘要】 地面自主移动机器人是一种能够在各种地面环境中连续自主运动的车辆,其发展对国防、社会、经济和科学技术具有重大的影响力,已成为各国高科技领域的战略性研究目标。而自主导航是自主移动机器人的一个最基本需求。同时定位与地图创建方法(SLAM)允许机器人在未知环境中,依靠自身所带的传感器递增式地创建环境地图,并同时给出机器人所在位置。自上世纪九十年代开始,该方法就吸引了国内外大量的研究者,并由于其重要的理论与应用价值被很多学者认为是实现真正全自主导航机器人的关键。近十年来,SLAM取得了令人瞩目的进展,并在室内、室外、水下、空中等多种环境下得到了实践。本文对移动机器人同时定位与地图创建方法进行了研究,在传统方法的基础上,提出了一些改进算法及新的解决方案,以提高SLAM算法的估计精度、一致性及计算效率,扩展其使用范围。具体的研究内容包括以下几个方面:1.从不确定性分析入手,在对SLAM中的相关性进行详细深入分析的基础上,得到了特征稀疏的两个标准,进而提出了相关优先的特征稀疏策略,从而减少大量的计算负担,计算误差却和一般传统方法相当。2.将SUT(scaled unscented transformation)变换运用到EKF SLAM算法中,研究了SLAM的线性化问题。3.针对Rao-Blackwellised粒子滤波SLAM(RBPF SLAM)算法的不一致现象,采用归一化估计方差(NEES)对算法的一致性进行了分析,得出粒子耗尽是造成算法不一致的原因,并分别采用辅助粒子滤波及正则粒子滤波对算法的重采样过程进行改进,提高了算法的一致性。4.针对普通粒子滤波容易受到粒子耗尽的影响,提出了一种新的粒子滤波SLAM算法。该算法将边缘粒子滤波技术(marginal particle filter,MPF)运用到SLAM中,并利用Unscented卡尔曼滤波(UKF)来计算提议分布。新算法避免了从不断增长的高维状态空间采样,非常有效地提高了算法中的有效粒子数,大大降低了粒子的权值方差,保证了粒子的多样性,同时也满足一致性要求。5.为了改善稀疏扩展信息滤波SLAM的性能,结合相关性分析,提出了一种改进的稀疏规则:完备信息稀疏规则。该稀疏规则考虑了预测时刻的观测信息,保留了与机器人相关性最强的主动特征。在不增加计算负担的情况下,提高了算法的精度及一致性。并对各种稀疏规则进行了深入的分析,就其优劣进行了比较,提出了一种组合的稀疏规则,以综合各自的优势,扩展使用环境范围。6.对SLAM中的联合数据关联方法进行了研究,提出了一种快速联合相容分枝定界算法。该方法通过给定每次联合相容配对个数的上限,来达到在观测数目较大时减少计算量的目的。当观测个数大于给定上限时,将分批进行数据关联,然后把结果组合起来。采用这种方法后,关联结果与一般的联合相容分枝定界算法差别很小,但计算量却大大降低。7.对SLAM问题中的重定位方法进行了研究,提出了一种改进的随机采样重定位方法。随机采样重定位的搜索部分的计算复杂度与观测数目成指数关系,当观测数目较大时,可通过快速联合相容分枝定界思想来达到减少计算量的目的。同时,在重定位算法中,为了防止误关联,一般会给定一个关联配对数的界限,只有配对数大于该界限时,才认为重定位是可靠的。但是,在很多情况下得到的配对数是小于给定界限的。通过引入运动约束来检验配对数较小时重定位的可靠性,以决定是否信任此时的重定位结果,达到提高重定位算法可用性的目的。本文在最后一章对全文进行了总结,并且对今后进一步的研究方向进行了展望。更多还原

【Abstract】 Autonomous Ground Vehicle (AGV) is an intelligent mobile robot, which can run autonomously, and continuously in real-time indoor or outdoor. The development of AGV has imposing on the defense, society, economy and academy, and becomes the tactic research object of high technology of all countries. Autonomous navigation is a fundamental problem for Autonomous Ground Vehicle.The simultaneous localization and map building (SLAM) problem asks if it is possible for a mobile robot to be placed at an unknown location in an unknown environment and for the robot to incrementally build a consistent map of this environment while simultaneously determining its location within this map. The SLAM problem has attracted a lot of researchers with a broad rang of interests and applications since 1990s. A solution to the SLAM problem has been seen as a "holy grail" for the mobile robotics community as it would provide the means to make a robot truly autonomous. The past decade has seen rapid and exciting progress in solving the SLAM problem together with many compelling implementation of SLAM methods. SLAM has been implemented in a number of different domains from indoor robots, to outdoor, underwater and airborne systems.This dissertation is focused on the SLAM problem. Several improved methods and novel solutions are presented in order to improve consistency and computational efficiency, and additionally extend SLAM application domains. The main content of this dissertation include the following aspects:1. Through the analysis of uncertainty, the correlation in SLAM problem is studied. It’s well-known that the correlation between features is actually the critical part of the SLAM problem. Maintaining and renewing this correlation information brings a huge computation burden. Furthermore, on the basis of having carried out deep analysis on correlation, a new feature sparse tactic named correlation priority is brought forward, which may use less features which having strong correlation to cut down large amount of the computation burden, and the computation error of this method can compare with that of some general traditional methods.2. In SLAM problem, motion and measurement models are usually of a very high nonlinear nature. An approach is designed which aims to avoid the analytical linearization based on Taylor-series expansion of both motion and measurement models by using scaled unscented transformation.3. The inconsistency problem of the Rao-Blackwellised particle filter (RBPF) SLAM algorithm is analyzed by using the normalized estimation error square (NEES). The result shows that it is sample impoverishment of particle filter which cause the inconsistency. So it is necessary to reduce the impact of resampling. Auxiliary particle filter and regularized particle filter are used to improve the RBPF SLAM resampling step in order to obtain consistent RBPF SLAM.4. In order to avoid the sample impoverishment problem of particle filter, a new particle filter SLAM algorithm is proposed, which is based on the marginal particle filter and using unscented Kalman filter (UKF) to generate proposal distributions. The underlying algorithm operates directly on the marginal distribution, hence avoiding having to perform importance sampling on a space of growing dimension. Additionally, UKF can reduce linearization error and gain accurate proposal distributions. Compare with the conventional particle filter SLAM methods, the new algorithm increases the number of effective particles and reduces variance of particles weight effectively. Also, it is consistent owing to the better particle diversity. As a result, it does not suffer from some shortcomings of existing particle methods for SLAM and has distinct superiority.5. In order to enhance the performance of sparse extended information filter (SEIF) SLAM algorithm, a novel sparsification rule is brought forward. The rule takes into account the observation information of sparsification time. The correlation can be observed globally and the features which have the strongest correlation are reserved. Therefore, the new algorithm gets higher estimation precision and more consistent results than the conventional SEIF algorithm by increasing no computation burden. Furthermore, an integrated sparsification rale is brought forward. The application environments are expanded.6. A new data association algorithm named fast JCBB (FJCBB) is proposed by giving a bound of joint compatibility pairings based on joint compatibility branch and bound (JCBB) algorithm. FJCBB has the same association performance with JCBB. Furthermore, FJCBB’s computational cost increases very slowly as the number of observations increases, this character makes FJCBB more feasible than JCBB. When the number of observations is large, FJCBB will also be adequately fast for real time implementation.7. The relocation problem is studied and an improved random sampling (RS) relocation algorithm is proposed. In RS algorithm the search part is a component exponential in the number of observations. When the number of observations is large, its computational cost increases rapidly. Therefore, the algorithm is improved by using FJCBB algorithm. The improved algorithm’s search part is linear time proportional to the number of observations. Furthermore, generally an empirical threshold of associated pairings is used to prevent false positives in relocation problem. When the number of pairings is lower than the threshold, a new method named motion constrains is brought forward to verify the relocation reliability. If it’s reliable, the relocation results will be accepted.A summary of the research conclusions and a discussion on the most promising paths of future research work are presented in the last chapter of this dissertation.更多还原