A novel qualitative motion model based probabilistic indoor global localization method
Introduction
The goal of mobile robot localization is to estimate a robot pose with respect to a map of its environment. It is a basic requirement for the mobile robot to implement given tasks. Local pose tracking and global localization are two paradigms of robot localization [19], [24]. Local pose tracking involves robot pose tracking with an initial state estimation. Meanwhile, global localization focuses on estimating the global pose using only sensor data without knowing the initial pose. It provides the robot with the ability to deal with initialization at start-up and recovery in case of local pose tracking failure [17].
The general approach to address the global localization problem is to match the current observed surroundings with a prior global map. Typically, the matching procedure is implemented by measuring similarities between the sensor data and virtual observations of hypothetical poses. The performance of global localization is determined by the hypotheses generating strategy. Current global localization methods can be classified into two approaches according to their hypotheses generating strategies: probabilistic filter-based and heuristic searching-based. In probabilistic filter-based methods [2], [3], [7], [24], the hypotheses are generated randomly. Then, the conditional probabilistic distribution is approximated and updated in a discrete way over the state space of a robot pose. In heuristic searching-based methods [13], [15], [17], robot pose estimation is formulated as an optimization problem and the hypotheses are generated according to the heuristic information.
These methods have inherent weaknesses for mobile robot practical navigation. A robot using a probabilistic filter-based global localization method should generate as many multiple pose hypotheses as possible to increase the success rate of pose recovery. However, under conditions of motion model uncertainty, pose hypotheses could converge on an incorrect location if there are too few particles to cover the correct area [17]. The performance of heuristic searching-based methods is heavily dependent on the convergence rate, which is affected by many factors.
In this paper, we propose a novel global localization method. The proposed method aims to overcome the previous weakness described above. Without using the raw grid map, we represent the map with a set of observations associated to the hypothetical poses offline. With this map representation, we select particles from the hypothetical poses during the online localization. The candidate hypotheses sets are constrained by the odometry error bounds. Then the particle sets are obtained through a sampling process in the candidate hypotheses sets. To obtain a global consistent trajectory estimation, a hypotheses tracking process is implemented. The particles are able to converge on the correct location since the motion uncertainty is well considered.
The contributions of this study include three aspects as follow. First, the global grid map is represented in a localization-oriented way. The global localization efficiency can be improved as the virtual observations are computed offline. Second, instead of sampling and updating particles in the whole state space, the proposed enhanced particle filter employs a qualitative motion model to constrain the possible areas. In these areas, the samples with high weights are selected as particles. With this mechanism, we do not maintain large scale particles to track the robot pose. Furthermore, the accumulative error is eliminated. Third, we present a scheme that the global localization accuracy can be predesigned as expected. In the proposed method, the increasing of pose estimation precision has small effect on the localization efficiency.
Section snippets
Related work
Global localization on 2D grid map is a classic yet still popular research subject. We briefly review some significant researches on the two major categories: probabilistic filter-based and heuristic searching-based methods.
Probabilistic filter-based methods aim to compute and evaluate a probabilistic distribution over the space of pose states. The Bayes filter algorithm provides a general approach to compute the belief distribution [17]. Grid-based Markov localization [2], [7] and Monte Carlo
Method
In this section, the proposed global localization method is presented. We start by introducing the global grid map representing scheme. We then propose the global localization formulation based on the proposed map representation. We subsequently present the global localization process with the enhanced particle filter. Finally, the overall global localization algorithm is given.
Global localization simulations
To verify the effectiveness of the proposed method, we conduct several simulations using the public grid map of the Intel Research Laboratory in Seattle [32]. The robot trajectory is manually constructed. Thus we are able to evaluate the accuracy of these methods quantitatively.
For the proposed global localization method, we apply two hypothetical pose sample sets with the interval of 10 grids and 2 grids. The low-density sample set is used for coarse and fast localization. The high-density
Conclusions and future work
In this paper, a novel probabilistic global localization method is proposed and verified. The method presents a modified version of particle filter. The framework of the proposed global localization method is designed to ensure that the robot poses are tracked with high efficiency and robustness. With the proposed global map representation, less particles are required to cover the robot pose state space, which makes it feasible to improve efficiency. By sampling in candidate pose set, the
Acknowledgments
This work is supported by the National Natural Science Fund of China (Grant No. 6137079) and the Science Foundation for Youths of Anhui Province(Grant No. 1708085QF159).
Authors’ contributions
Jikai Wang and Peng Wang developed the method, carried out the experiments and wrote the manuscript. Zonghai Chen gave the research direction and ideas, and put forward amendments and improvements to the manuscript.
Ji-Kai Wang received the B.S. degree from the University of Yanshan in 2014. He is now a PhD in the Department of Automation, USTC. His research interests include small sample processing, mobile robots SLAM and knowledge representation.
References (32)
- et al.
A novel visual codebook model based on fuzzy geometry for large-scale image classification
Pattern Recogn.
(2015) - et al.
Efficient visual tracking using particle filter with incremental likelihood calculation
Inf. Sci.
(2012) - et al.
A novel efficient algorithm for mobile robot localization
Rob. Auton. Syst.
(2013) - et al.
Evolutionary filter for robust mobile robot global localization
Rob. Auton. Syst.
(2006) - et al.
Vision-based global localization for mobile robots with hybrid maps of objects and spatial layouts
Inf. Sci.
(2009) - et al.
Robust Monte Carlo localization for mobile robots
Artificial Intell.
(2001) - et al.
Segmentation and recognition of multi-model photo event
Neurocomputing
(2016) - et al.
Object tracking from image sequences using adaptive models in fuzzy particle filter
Inf. Sci.
(2013) - et al.
Corrective gradient refinement for mobile robot localization
- et al.
Estimating the absolute position of a mobile robot using position probability grids
Monte carlo localization for mobile robots
Efficient multi-robot localization based on monte carlo approximation
Int. Symp. Rob. Res.
Improving grid-based slam with rao-blackwellized particle filters by adaptive proposals and selective resampling
An experimental comparison of localization methods continued
Experiments on augmenting condensation for mobile robot localization
Adaptive real-time particle filters for robot localization
Cited by (0)
Ji-Kai Wang received the B.S. degree from the University of Yanshan in 2014. He is now a PhD in the Department of Automation, USTC. His research interests include small sample processing, mobile robots SLAM and knowledge representation.
Peng Wang received the B.S. degree from the University of Science and Technology of China (USTC) in 2010. He is now a post doctor in the Department of Automation, USTC. His research interests include system modeling and simulation, SLAM and deep learning.
Zong-Hai Chen received the B.S. degree from the University of Science and Technology of China (USTC) in 1988. He is currently a professor at the Laboratory of Simulation and Intelligent Control in the Department of Automation, USTC. His research interests include modeling and simulation of complex system, control system engineering and intelligent information processing. Prof. Chen is a member of the Robotics Technical Committee of the International Federation of Automation Control (IFAC).