Improved Techniques for Grid Mapping With Rao-Blackwellized Particle Filters全文翻译Gmapping
Improved Techniques for Grid Mapping With Rao-Blackwellized Particle Filters
Abstract—Recently, Rao-Blackwellized particle filters (RBPF)have been introduced as an effective means to solve the simultaneous localization and mapping problem. This approach uses a particle filter in which each particle carries an individual map of the environment. Accordingly, a key question is how to reduce the number of particles. In this paper, we present adaptive techniques for reducing this number in a RBPF for learning grid maps. We propose an approach to compute an accurate proposal distribution, taking into account not only the movement of the robot, but also the most recent observation. This drastically decreases the uncertainty about the robot’s pose in the prediction step of the filter. Furthermore, we present an approach to selectively carry out resampling operations, which seriously reduces the problem of particle depletion. Experimental results carried out with real mobile robots in large-scale indoor, as well as outdoor, environments illustrate the advantages of our methods over previous approaches.
近年来,Rao-Blackwellized粒子滤波器(RBPF)被引入到解决同时定位和建图问题中。此方法使用粒子过滤器,其中每个粒子都携带环境的单个地图。因此,一个关键问题是如何减少粒子的数量。在这篇文章中,我们提出了在学习网格地图的RBPF中减少粒子数量的自适应技术。我们提出了一种计算精确建议分布的方法,不仅考虑了机器人的运动,而且还考虑了最近的观察。这大大降低了滤波器建议步骤中机器人姿态的不确定性。此外,我们还提出了一种选择性地执行重采样操作的方法,这大大减少了粒子耗尽的问题。在大型室内和室外环境中,用真实的移动机器人进行的实验结果说明了我们的方法比以前的方法的优势。
Index Terms—Adaptive resampling, improved proposal, motion model, Rao-Blackwellized particle filter (RBPF), simultaneous localization and mapping (SLAM).
索引项-自适应重采样、改进方案、运动模型、Rao-Blackwellized粒子滤波器(RBPF)、同步定位与建图(SLAM)。
I. INTRODUCTION
Building maps is one of the fundamental tasks of mobile robots. In the literature, the mobile-robot mapping problem is often referred to as the simultaneous localization and mapping (SLAM) problem [1]–[9]. It is considered to be a complex problem, because for localization, a robot needs a consistent map, and for acquiring a map, a robot requires a good estimate of its location. This mutual dependency between the pose and the map estimates makes the SLAM problem hard, and requires searching for a solution in a high-dimensional space.
地图绘制是移动机器人的基本任务之一。在文献中,移动机器人的建图问题通常被称为同时定位和映射(SLAM)问题[1]-[9]。这被认为是一个复杂的问题,因为对于定位,机器人需要一个一致的地图,而对于获取地图,机器人需要对其位置进行良好的估计。姿态估计和地图估计之间的这种相互依赖使得SLAM问题变得困难,需要在高维空间中寻找解决方案。
Murphy, Doucet, and colleagues [2], [8] introduced Rao-Blackwellized particle filters (RBPF) as an effective means to solve the SLAM problem. The main problem of the Rao-Blackwellized approaches is their complexity, measured in terms of the number of particles required to build an accurate map. Therefore, reducing this quantity is one of the major challenges for this family of algorithms. Additionally, the resampling step can potentially eliminate the correct particle.This effect is also known as the particle-depletion problem, or particle impoverishment [10].
Murphy,Doucet及其同事[2],[8]介绍了Rao-Blackwellized粒子滤波器(RBPF)作为解决SLAM问题的有效方法。Rao-Blackwellized方法的主要问题是它们的复杂性,用建立精确地图所需的粒子数来衡量。因此,减少粒子数量是这类算法面临的主要挑战之一。此外,重采样步骤可以潜在地消除正确的粒子。这种效果也称为粒子耗尽问题,或粒子贫化[10]。
In this paper, we present two approaches to substantially increase the performance of RBPFs applied to solve the SLAM problem with grid maps:
•a proposal distribution that considers the accuracy of the robot’s sensors and allows us to draw particles in a highly accurate manner;
• an adaptive resampling technique which maintains a reasonable variety of particles, and in this way, enables the algorithm to learn an accurate map while reducing the risk of particle depletion.
在本文中,我们提出了两种方法来显著提高RBPFs的性能,并将其应用于解决栅格地图的SLAM问题:
•一种建议分布,不仅考虑到机器人传感器精度并允许我们以高精度方式绘制粒子;
•一种自适应重采样技术,它保持了粒子的合理多样性,并以这种方式使算法能够学习精确的映射,同时降低粒子耗尽的风险。
The proposal distribution is computed by evaluating the likelihood around a particle-dependent most-likely pose, obtained by a scan-matching procedure combined with odometry information. In this way, the most recent sensor observation is taken into account for creating the next generation of particles. This allows us to estimate the state of the robot according to a more informed (and thus, more accurate) model than the one obtained based only on the odometry information. The use of this refined model has two effects. The map is more accurate, since the current observation is incorporated into the individual maps after considering its effect on the pose of the robot. This significantly reduces the estimation error, so that fewer particles are required to represent the posterior. The second approach, the adaptive resampling strategy, allows us to perform a resampling step only when needed, and in this way, keeps a reasonable particle diversity. This results in a significantly reduced risk of particle depletion.
建议分布是通过评估与粒子相关的最可能姿态周围的可能性来计算的,该可能性是通过扫描匹配过程与里程信息相结合获得的。以这种方式,最新的传感器观测被考虑到创造下一代粒子。这使得我们能够根据一个比仅基于里程信息获得的模型更为富含信息(从而更为准确)的模型来估计机器人的状态。使用这种改进的模型有两个效果。考虑到当前的观测结果对机器人姿态的影响,该地图更加精确。这显著地减少了估计误差,因此需要更少的粒子来表示后验。第二种方法,自适应重采样策略,允许我们仅在需要时执行重采样步骤,这样,就保持了合理的粒子多样性。这将显著降低微粒耗尽的风险。
The work presented in this paper is an extension of our previous work [11], as it further optimizes the proposal distribution to even more accurately draw the next generation of particles.Furthermore, we added a complexity analysis, a formal description of the techniques used, and provide more detailed experiments in this paper. Our approach has been validated by a set of systematic experiments in large-scale indoor and outdoor environments. In all experiments, our approach generated highly accurate metric maps. Additionally, the number of the required particles is one order of magnitude lower than with previous approaches.
本文的工作是我们先前工作的扩展[11],因为它进一步优化了建议分布,以更准确地绘制下一代粒子。此外,我们还添加了复杂性分析,对所使用的技术进行了形式化描述,并在本文中提供了更详细的实验。我们的方法已经在一系列大型室内外环境的系统实验中得到了验证。在所有实验中,我们的方法生成了高精度的度量地图。此外,所需粒子的数量比以前的方法低一个数量级。
This paper is organized as follows. After explaining how a Rao-Blackwellized filter can be used in general to solve the SLAM problem, we describe our approach in Section III. We then provide implementation details in Section IV. Experiments carried out on real robots are presented in Section VI. Finally,Section VII discusses related approaches.
本文的结构如下。在解释了如何使用Rao-Blackwellized滤波器来解决SLAM问题之后,我们在第三节描述了我们的方法。然后在第四节提供了实现细节。在第六节中介绍了在真实机器人上进行的实验。最后,第七节讨论了相关的方法。
II. MAPPING WITH RBPFS
According to Murphy [8], the key idea of the RBPF for SLAM is to estimate the joint posterior p(x_(1:t),m|z_(1:t),u_(1:t-1))about the map mand the trajectory x_(1:t)=x_1,…,x_tof the robot. This estimation is performed given the observations z_(1:t)=z_1,…,z_tand the odometry measurements u_(1:t-1)=u,…,u_(t-1)obtained by the mobile robot. The RBPF for SLAM makes use of the following factorization:
根据Murphy[8],SLAM的RBPF的关键思想是估计地图上的联合后验概率p(x_(1:t),m|z_(1:t),u_(1:t-1))和机器人的轨迹x_(1:t)=x_1,…,x_t。在给定移动机器人获得的观测值z_(1:t)=z_1,…,z_t和里程测量值u_(1:t-1)=u,…,u_(t-1)的情况下,进行该估计。SLAM的RBPF使用以下因子分解:
p(x_(1:t),m|z_(1:t),u_(1:t-1))= p(m│x_(1:t),z_(1:t) )·p(x_(1:t)│z_(1:t),u_(1:t-1) ) (1)
This factorization allows us to first estimate only the trajectory of the robot, and then to compute the map given that trajectory. Since the map strongly depends on the pose estimate of the robot, this approach offers an efficient computation. This technique is often referred to as Rao-Blackwellization.
这个分解允许我们首先估计机器人的轨迹,然后计算给定轨迹的地图。由于地图高度依赖于机器人的姿态估计,这种方法提供了一种有效的计算方法。这种技术通常被称为Rao-Blackwellization。
Typically, (1) can be calculated efficiently, since the posterior over maps p(m│x_(1:t),z_(1:t) ) can be computed analytically using “mapping with known poses” [12], since x_(1:t)andz_(1:t) are known.
通常,(1)可以被有效地计算,因为x_(1:t)和z_(1:t) 是已经知道的。地图的概率密度函数可以通过已知位姿的建图方法“mapping with known poses”来计算。
To estimate the posterior over the potential trajectories, one can apply a particle filter. Each particle represents a potential trajectory of the robot. Furthermore, an individual map is associated with each sample. The maps are built from the observations, and the trajectory represented by the corresponding particle.
为了估计潜在轨迹上的后验概率p(x_(1:t)│z_(1:t),u_(1:t-1) ),可以应用粒子滤波器。每个粒子代表机器人的一个潜在轨迹。此外,一个单独的地图与每个样本相关联。这些地图是根据观测结果建立的,轨迹由相应的粒子表示。
One of the most common particle-filtering algorithms is the sampling importance resampling (SIR) filter. A Rao-Blackwellized SIR filter for mapping incrementally processes the sensor observations and the odometry readings as they are available. It updates the set of samples that represents the posterior about the map and the trajectory of the vehicle. The process can be summarized by the following four steps.
最常用的粒子滤波算法之一是(SIR)滤波器。用于建图的Rao-Blackwellized SIR滤波器,增量式地处理传感器观测值和里程计读数(当他们可获取的时候)。它更新代表地图和车辆轨迹的后验概率的样本集。这个过程可以概括为以下四个步骤。
- Sampling: The next generation of particles {x_t^((i) )} is obtained from the generation {x_(t-1)^((i) ) }by sampling from the proposal distribution π. Often, a probabilistic odometry motion model is used as the proposal distribution.
- Importance Weighting: An individual importance weight{w_t^((i) )} is assigned to each particle, according to the importance sampling principle
The weights account for the fact that the proposal distribution is, in general, not equal to the target distribution of successor states.
3) Resampling: Particles are drawn with replacement proportional to their importance weight. This step is necessary,since only a finite number of particles is used to approximate a continuous distribution. Furthermore, resampling allows us to apply a particle filter in situations in which the target distribution differs from the proposal. After resampling, all the particles have the same weight.
4) Map Estimation: For each particle, the corresponding map estimate p(m^i |x_(1:t)^((i) ),z_(1:t)) is computed, based on the trajectory x_(1:t)^((i)) of that sample and the history of observations z_(1:t)。
1)采样:下一代粒子{x_t^((i) )}是从{x_(t-1)^((i) ) }一代中通过从建议分布π采样获得的。通常,建议分布使用概率里程运动模型。
2)重要性权重:根据重要性抽样原则,为每个粒子分配一个单独的重要性权重{w_t^((i) )}
这些权重说明了这样一个事实,即建议分布一般不等于继承状态的目标分布。
3) 重采样:根据权值的比例重新分布采样粒子。这一步是必要的,因为只有有限数量的粒子用于近似连续分布。此外,重采样允许我们在目标分布与建议分布不同的情况下应用粒子过滤器。重采样后,所有粒子具有相同的权重。
4) 地图估计:对于每个粒子,根据该样本的轨迹x_(1:t)^((i))和观测历史z_(1:t),计算相应的Map估计 p(m^i |x_(1:t)^((i) ),z_(1:t))
The implementation of this schema requires evaluating the weights of the trajectories from scratch whenever a new observation is available. Since the length of the trajectory increases over time, this procedure would lead to an obviously inefficient algorithm. According to Doucet et al. [13], we obtain a recursive formulation to compute the importance weights by restricting the proposal to fulfill the following assumption:
这个方案的实现需要在获取新的观测时从头开始评估轨迹的权重。由于轨迹的长度随着时间的推移而增加,此过程将导致明显低效的算法。根据Doucet等人的说法。[13] ,通过限制建议分布满足以下假设,我们获得了计算重要性权重的递归公式:
Here μ=1⁄(p(z_t |z_(1:t-1),u_(1:t-1))) is a normalization factor resulting from Bayes’ rule that is equal for all particles.
这里μ=1⁄(p(z_t |z_(1:t-1),u_(1:t-1)))是由Bayes规则产生的标准化因子,对所有粒子来说是一样的。
Most of the existing particle filter applications rely on the recursive structure of (6). Whereas the generic algorithm specifies a framework that can be used for learning maps, it leaves open how the proposal distribution should be computed and when the resampling step should be carried out. Throughout the remainder of this paper, we describe a technique that computes an accurate proposal distribution, and that adaptively performs resampling.
现有的粒子滤波应用大多依赖于(6)的递归结构。尽管通用算法指定了一个可用于学习地图的框架,但它对如何计算建议分布以及何时执行重采样步骤仍有开放性。在本文的其余部分中,我们描述了一种计算精确建议分布并自适应执行重采样的技术。
III. RBPF WITH IMPROVED PROPOSALSAND ADAPTIVE RESAMPLING
In the literature, several methods for computing improved proposal distributions and for reducing the risk of particle depletion have been proposed [13]–[15]. Our approach applies two concepts that have previously been identified as key prerequisites for efficient particle-filter implementations (see Doucet et al. [13]), namely, the computation of an improved proposal distribution and an adaptive resampling technique.
在文献中,已经提出了几种计算改进的建议分布和降低粒子耗尽风险的方法[13]-[15]。我们的方法应用了两个概念,这两个概念以前被认为是高效粒子滤波器实现的关键先决条件(参见Doucet等人。[13] ,即计算改进的建议分布和自适应重采样技术。
A. On the Improved Proposal Distribution
As described in Section II, one needs