Localizability Estimation And Localization For Mobile Robots

Localization for mobile robots is one of the key technologies in robotics research area,which generally includes two main sorts: global localization and pose-tracking. A reliablerobot pose (position and heading) is necessary for mobile robots to complete differenttasks. But for global localization, the mobile robot needs to face the influences onlocalization from the prior-map (e.g. different map scenes with unspecific features andnoises), besides of the error from the sensors (e.g. odometer). These influences will slowthe convergence of the global localization and even lead to a failed convergence result.For pose-tracking, the mobile robot will suffer the greater influence on localization fromthe dynamic obstacles (e.g. moving people), besides of the above two kinds of influences.The combined influences will degrade the localization accuracy and even lead to lose thepose (i.e. kidnap problem) that, the mobile robot is unable to continue the tasks.Concerning with the global localization and pose-tracking technologies for mobilerobots, this study gives the detailed analyses and the estimation of the localizability ofmobile robots. Based on the localizability, the improved methods are proposed. As aresult, the convergence of global localization is accelerated and the robustness isenhanced; and also the localization accuracy of pose-tracking is upgraded and itsrobustness is enhanced. The main research points are as below:1. A method for estimation and quantization of robot localizability is proposed basedon the prior-map (probabilistic grid map, PGM) and observation (laser range-finder, LRF)models. Essentially, this method is based on the estimation of the localization covariancematrix accoriding to Fisher’s Information Matrix (FIM) and Cramér-Rao Bound (CRB).By discretization of FIM, the calculation of the localizability matrix in the environmentswithout dynamic obstacles is given. This matrix is used for the pre-cache of the robotlocalizability on the prior-map, which is defined as the static localizability matrix. Thenan impact factor of the dynamic obstacle is defined based on Bayesian algorithm.Combining the impact factor with the statice localizability matrix, the dynamiclocalizability matrix is given, which is used to reflect the influences on localization fromthe prior-map and dynamic obstacles on-line. Through analyses of the covariance ellipse derived from the localizability matrix, we can get that the localizability matrix canaccurately represent not only the respective but also the coupling influences onlocalization between the position and heading states.2. Concerning with global localization, to accelerate the convergence and enhancethe robustness in the different scenes with unspecific features and map noises, this studyproposed an active global localization algorithm based on the localizability. Specifically,the algorithm framework is based on the standard particle filtering (PF). According to theaction selection mechanism determined by the static localizability matrix, each selectedrobot action will make the whole particle system achieve the maximum localizabilitydistinctness, through which the convergence and robustness of the particl system isimproved. The theoretical analyses of the convergence, effectiveness and real-timing aredone. We can see that, the proposed algorithm considers the influences on globallocalization from the different map scenes and noises, without extraction of any specificfeatures. Additionally, comparing with the passive algorithm in which the robot actionsare random, as the static localizability matrix is pre-cached, the computational complexityof the proposed algorithm does not increase greatly.Based on the open CARMEN toolkit for the robot operation system, the proposedlocalizability-based active global localization algorithm is compared respectively with theclassical passive and active algorithms in simulations. Then the proposed algorithm iscompared with the passive algorithm through experiments in the real environments (theoffice and corridor), using “JiaoLong” intelligent wheelchair which equips the odometerand LRF. The experimental results of the proposed algorithm are also compared with theexisting classical active algorithms. The simulation and experimental results show that,the proposed active global localization algorithm is effective to accelerate theconvergence and enhance the robustness of the global localization with sensor error in thedifferent scenes with unspecific features and map noises.3. Concerning with pose-tracking, to maintain the localization accuracy androbutstness in the different environments under the high and dynamic occlusions, thisstudy proposed a self-adaptive pose-tracking algorithm based on the localizability.Concretely, under the framework of the standard PF, the odometer-based proposaldistribution function (PDF) is corrected using the dynamic localizability matrix estimatedaccording to the LRF observation model. To guarantee the robustness against the differentocclusion degrees, the reliability parameter between the measurements of odometer andobservations of LRF is improved. The convergence, effectiveness and real-timing of the algorithm are verified through the theoretical analyses. We can see that, when thelocalizability is weak, PDF is mainly decided by the odometer process model, whichguarantees the huge fluctuation of localization results; and when the localizabilitybecomes strong, the observation model becomes important, which guarantees theimmediate correction of PDF. Additionally, as the improved PDF is more accurate evenusing fewer particles, the computational time is ensured.Based on the CARMEN toolkit, the proposed localizability-based self-adaptivepose-tracking algorithm is compared with the classical algorithms in simulations. Thenusing “JiaoLong” intelligent wheelchair, the proposed algorithm is compared with theclassical algorithms through experiments in the real and crowded environments (the office,corridor, cafeteria and metro station). The simulation and experimental results show that,the proposed self-adaptive pose-tracking algorithm maintains an accurate and robust robotpose in the different scenes with sensor error and map noises against the high anddynamic occlusions.In summary, the effectiveness of the proposed active global localization and theself-adaptive pose-tracking algorithms illustrates that, the defined localizability can beused to improve the localization technologies of mobile robots, and also the informationfusion method is valid. These lay a solid foundation for the future real application ofmobile robots in the daily life.