On Recovering Localization For SLAM

As an indispensable part for autonomous mobile robots,it has been widely investigated on retrieving precisely the location and orientation after kidnapping without any priori knowledge about the new position,whitch is called localization recovery.However,current localization recovery technology still has some challenging problems that have not been competely adressed,such as poor reliability,slow speed and inability to cope with dynamic environment.In order to solve these problems,this thesis mainly studies localization recovery technique of mobile robots,and proposes two localization recovery solutions for the existing laser SLAM(simultaneous localization and mapping)technology.Firstly,a feature-based localization recovery solution is developed in this thesis.This solution constructs an occupancy grid map with the help of SLAM technology.Next,the hidden key points are obtained by extracting straight lines in the occupancy grid map and calculating the intersection points of straight lines.Then,LBP(local binary pattern)operators are adopted to describe the key points to reduce the error matching.In the meanwhile,multiple groups of key point pairs are determined by using the geometric matching method.For each group of key point pairs,a rough pose is obtained using the random sample consensus(RANSAC)algorithm.Finally,the local point-cloud based map is compared with the global one to estimate the robot pose together with an evaluation value for each estimation,and the robot pose with the highest value is selected as the final estimated pose.This solution is highly applicable,which improves the reliability of localization recovery and can be implemented on the on-board controller.Secondly,a localization recovery solution based on a neural network is proposed.Localization recovery model is trained offline through collected laser data and the corresponding pose obtained from SLAM.From the model,a rough pose estimation can be obtained,based on which,a precise pose estimation is then generated through point cloud based map matching with the global map.Moreover,a motion control scheme is proposed to obtain the final estimated pose by combining multiple localization recovery during the motion.This solution avoids the complicated feature extraction process,and can realize fast localization recovery online through the offline learning and training model.Meanwhile,it has high reliability in dynamic environment.