| Mobile robot global localization is one of the basic and important modules in the navigation system,which determines the performance of navigation system under dif-ferent environmental conditions.To implement a given task,the robot should be able to estimate its pose accurately and steadily.Mobile robots estimate their pose by match-ing current observations and historical observations with existing environment maps.However,there are many factors in the process,such as sensor noise,environmental dynamics,ambiguity and the incompleteness of environmental perception information,which affect the mobile robot global localization due to multi-source uncertainty.These uncertainties include the uncertainty of observation and the uncertainty of feature loca-tion association.The observation information deviates from the real physical informa-tion because of environmental conditions,sensor noise and information transmission noise.On account of insufficiency of robot's perception and cognitive ability to the en-vironment,there is uncertainty in the global position estimation of currently observed features to the robot.Both of these uncertainties affect the robot's position estimation.The influence of the latter leads to the uncertainty of the robot's location in the environ-ment.The former mainly leads to the uncertainty of the robot's precise pose.Therefore,this dissertation studies the uncertain information modeling method for the localization task,and then builds the environment information storage,retrieval,information asso-ciation and reasoning methods.It is very important to improve the performance of the global location system for mobile robot.This dissertation leverages the advantages of the qualitative reasoning,probability theory,machine learning and other representation and processing methods of uncertain information to model the uncertainty of observa-tion and the uncertainty of feature location association.Some novel global localization methods are proposed to improve the stability,efficiency and intelligence of mobile robot autonomous navigation.In order to improve the processing ability of the information containing observation uncertainty,it is necessary to establish the information representation and processing model that is suitable for the characteristics of the information.Based on the theory of qualitative reasoning and probability,a qualitative particle filter is proposed to improve the performance of laser global localization for mobile robot.In order to improve the ability to deal with the uncertainty of feature association,it is necessary to improve the performance of feature representation and association model,and the performance of feature association reasoning methods.In this thesis,a self-learning method for feature-location mapping relationship of geometric map,point feature map,line feature map and topological map under the supervision of localization task is established based on regression forest model.And a feature association probabilistic reasoning mechanism is proposed based on an improved RANSAC framework.The main jobs and innovations of this thesis are as follows:1)In the view of the uncertainty of the state probability distribution caused by the observation noise and the uncertainty of the correlation due to the low representation ability of the laser in the process of laser global localization,a qualitative particle filter model is established by combining the qualitative reasoning method and the Bayesian method.And a global localization method is proposed based on the model.Firstly,the environment map is represented as a set of qualitative particles and their expected observations,and the set of mapping relationship between the position and observation in the environment is pre-constructed,which avoids the calculation of expected obser-vations online.In the process of online state iterative estimation,due to the parameters uncertainty of state transition model and observation model,qualitative state transition model and qualitative observation model are proposed.The complete estimation and coverage to the possible distribution area of the robot's pose are realized based on the qualitative model,and the real state distribution of the robot is estimated by the parti-cles contained in the possible area.Finally,the probability transfer model of context is established to realize the convergence of state estimation from multiple hypotheses to global optimum.The experimental results show that the proposed method is robust to the modeling errors of the state transition model.At the same time,it is computationally efficient because it can adaptively adjust the particles' size.2)In the view of the uncertainty of camera pose estimation caused by the weak-ness of visual feature representation ability and the existence of environmental repeated texture and structure in the process of visual global localization,this thesis proposes a visual global localization method based on regression forest.This method includes offline self-learning of environment model and online camera pose estimation.First of all,self-learning model of environment topology structure is realized by regression tree,which establishes the topology regression tree to split the local scenes.Furthermore,the regression forest is used to model and manage the uncertainty of the relationship between the visual point features and the geometric spatial position.In the process of online localization,the robot is first topologically located by using topological regres-sion tree,and then the fine localization is performed by using feature point regression tree in topological nodes.This hierarchical global localization strategy greatly improves the efficiency of geometric reasoning in the localization process.On this basis,a graph model based representation method of geometric consistency information and the corre-sponding association reasoning method are proposed to optimize the position prediction of feature space.The experimental results in the public dataset show that the proposed method can achieve high global positioning accuracy in indoor scenes.3)In order to further improve the adaptability of the visual global localization method in different scenes,a visual global localization method is proposed based on multiple geometric information of environments in this thesis.Firstly,learning based method for feature-location mapping relationship of geometric map,point feature map,and line feature map is established based on regression forest.The point feature regres-sion tree and line feature regression tree can predict the spatial position of the feature given the observed point feature and line feature.At the same time,in the localization process,in order to retrieve the geometric surface of the environment surface quickly,a compressed representation model of environment dense map is proposed based on regression tree,which greatly improves the efficiency and accuracy of pose hypothesis selection.During the online localization process,the three-dimensional spatial posi-tions of the currently observed point features and line segment features are predicted by the learned model,and then the camera pose inference is carried out according to the predicted positions.In this thesis,a method of pose calculation based on points and line segments is proposed.In the process of generating pose hypotheses,a stacked-RANSAC is proposed to improve the probabilistic reasoning performance of feature as-sociation.The results on public dataset and online experiments show that the proposed method achieves high global localization performance in multiple textured scenes. |
PDF Full Text Request