Research On Robust Visual SLAM And Autonomous Navigation Systems For Indoor Mobile Robots

Localization and Mapping are two key problems to be solved in autonomous navigation for mobile robots.Additionally,the estimation precision and robustness of an algorithm will have a significant influence on the path planning and motion control of any given robot.Simultaneous Localization and Mapping(SLAM)is an important technology towards addressing these two key problems,and has a wide range of applications.This has transformed SLAM into one of the research hotspots in the field of robotics in recent years.Furthermore,due to its low implementation cost and additional ability to exploit higher dimensions and richer information from the visual sensor,Visual SLAM has become one of the most important research fields of SLAM.This thesis mainly aims at solving the shortcomings of the traditional Visual SLAM algorithm based on feature points matching.This work proposes some improvement methods towards enhancing localization robustness and accuracy.Finally,a set of autonomous navigation systems for an indoor mobile robot,based on Visual SLAM is implemented.The system is shown to overcome the route rigidity and poor flexibility problems associated with traditional navigation technologies such as electro-magnetic navigation.The main research work and contributions include:? A introduction to the background and significance of Visual SLAM,as well as the navigation methods commonly used in mobile robotics.A comprehensive introduction to the background principles and theory of Visual SLAM,including the models of the various cameras,the pose description based on Lie group and Lie algebra,as well as the system framework based on Graph Optimization.? An analysis of the multiple local minima problem associated with traditional Bundle Adjustment(BA)optimization methods based on re-projection errors.We put forward an appropriate modification to the re-projection error.A BA optimization method based on polar line segmentation is proposed.The proposed method adds the 2D-2D epipolar constraints based on the 3D-2D constrains.Finally,we apply the proposed method to an ORB-SLAM2 system and conduct experimental verification.Results show that the robustness and accuracy of the system can be improved through the proposed method.? We address the lost problem associated with the ORB-SLAM2 algorithm which leads to navigation drawbacks.We also address an existing problem of Visual SLAM based on image feature points.This problem is associated with an efficient use of image information and a large computational load associated with the algorithm's point matching scheme.The sparse direct RGBD-DSLAM algorithm is proposed,which estimates the pose of the camera by minimizing the photometric error of the feature point image block between adjacent images.RGBD-DSLAM is shown to have the potential of improving the computational efficiency and robustness of localization.? Finally,a Visual SLAM-based autonomous navigation system for indoor mobile robots is implemented.This implementation covers the mobile robot hardware system as well as the software framework based on Robot Operating System(ROS).A three-dimensional(3D)sparse feature point map and a two-dimensional(2D)grid map are simultaneously constructed by using the robust Visual SLAM algorithm in an unknown environment.We validate the system's ability to avoid the problem of map pair misalignment.We achieve a comprehensive use of the improved ORB-SLAM2 and RGBD-DSLAM algorithms in realizing the mobile robot's localization within this autonomous navigation system.Navigation experiments show that the system has good robustness and accuracy.Finally,a method for rapid three dimensional(3D)dense mapping is also proposed and verified for the implemented robot system.