Research On Consistent Methods For Simultaneous Localization And Mapping Of Outdoor Mobile Robots

Simultaneous localization and mapping (SLAM) is a problem of joint stateestimation, which builds the environmental map incrementally and estimates itspose when a mobile robot is exploring an unknown environment. This thesismainly uses laser scanner as a range-bearing sensor to measure the relativepositions between the objects in the environment and the robot and to build thepoint feature map. Focusing on the problem of consistency for solving SLAM,some improved methods are put forward and applied to map building ofintelligent vehicle and automatic homing of scout robot.When using an extended Kalman filter (EKF) to solve SLAM, a fullcovariance matrix is utilized to maintain the cross correlations between differentcomponents of the joint state so that the invisible features could be corrected ateach state update step. But the accumulated errors caused by imprecise modelsand linearization make it difficult to keep the covariance matrix accurately,which results in wrong estimate of the state’s mean and produces optimisticestimation. This thesis first validates that the linearization errors will causeinconsistency certainly when the robot keeps stationary and observes newfeatures repeatedly. And the main factor is the big variance of the robot’sheading angle. In order to improve the consistency, a movable coordinate frame(MCF) based SLAM method is proposed, which shifts the reference frame to theplace of the robot’s current pose once a new feature is detected. So the initialvariance of each feature is only determined by the measurement errors and isirrelevant with the uncertainty of the robot pose. Three factors, i.e. the weakinitial correlations between components, the small Kalman gain of the robot’sheading angle and the boundedness of the pose’s uncertainty, provide the basisfor the movement of the coordinate frame. By using the improved method, therobot will explore a known environment all along and the inconsistency causedby linearization errors is limited in local environments. Since it has square computational complexity for EKF to maintain acovariance matrix, an extended information filter (EIF) is used to solve SLAM.This thesis uses a graphical model as the data structure and explains that thelocal performance of EIF could be preserved if the robot’s old pose is reservedwhen moving the coordinate frame and the new pose will has no connectionswith the old features unobserved. In order to have good consistency andcomputational performance concurrently, a MCF based sparse information filteris proposed, which moves the coordinate frame and reserves the robot posewhen the robot switches between local environments. So the information matrixis kept sparse exactly and the exploration in local environments has onlyconstant time complexity. If there are dense features in the environment, anexisting sparsification approximation is utilized to make the information matrixsparser.In order to realize large scale environment exploration, a constraint basedsubmap algorithm is proposed, which utilizes the correspondences between thecommon features in geometrically adjacent submaps and the relationshipsbetween the robot poses and the local refererence frames so that the weightedleast squares constraints between submaps are built. After closing the loop of theenvironment, an equation combining the constraints is calculated by usingiterative optimization. Then the global poses of the local reference frames areyielded, which are followed by the partial adjustment of the tail of each submap.At last, the globally consistent map is acquired.Finally, the problem of automatic homing for a scout robot in the process ofexploring dangerous environments is discussed. A way-point based matchingmethod is proposed, which uses the iterative closest point algorithm to graduallyadjust the returning pose coherent with the recorded way-points. When theenvironment is complicated, the original measurement data without extraction isused to represent the environment. The robot’s pose at each instance is one partof the vector and a big equation with weighted least squares constraints isformed. So a global map with the original representation is acquired, whichcould localize the robot at the returning stage.