Research On Visual Slam System Based On Multi-source Information Fusion

With the improvement in power and speed of computer processors,Visual Simutaneous Localization and Mapping(VSLAM),due to the decrease of camera size and cost,has gradually become a hot topic in robotics community and computer vision community in rencent years.Most of VSLAM systems that only use pure visual information are usually work well in experimental environments,but may decay in real-world environments with low-texture scenes and lighting changes.Meanwhile,those systems rarely construct environmental map with structural information.Due to the scale uncertainty and image blur in camera measurements when moving rapidly with the robot,those traditional VSLAM systems cannot actually be used for autonomous robot navigation.In order to solve these problems,this dissertation leverages multi-source information fusion technique to tightly fuse point and line features with Inertial Measurement Unit(IMU)measurements,and develops a VSLAM system based on non-linear graph optimization algorithm..At first,the basic dynamic model of the VSLAM system is introduced,and the related visual and inertial techniques used in motion estimation is also elaborated.Then,the VSLAM state estimation problem based on multi-source information fusion are analyzed mathematically.After measurement models o f camera and IMU are analyzed,the re-projection error of point and line features are built with the pre-integration error of IMU at the same time.The state estimation problem are finally solved by non-linear graph optimization algorithm based on the sliding window technique.This dissertation also introduces the Plücker Coordinates to parameterize line features,and improves the triangulation method in traditional VSLAM systems.In order to solve the problem of redundant degrees of freedom in the Plücker coordinate system and improve the positioning accuracy of the system,the Closest Point Representation is introduced to describe the line state during non-linear graph optimization.By leveraging Vins-Mono,the design and implementation of this VSLAM system based on multi-source information fusion is introduced.The proposed system mainly includes three modules: feature tracker,state estimation and loop-closure.The program flow,principles and details of each module are explained in detail.At the end,a simulation on the proposed line triangulation method are carried out,where the results show that the proposed method can effectively restore the line features in the environment,and improve the performance of the localization to a certain extent.The system is also verified in the real-world experiments using open EuRoC dataset.Compared with the traditional VSLAM systems,the results show that the proposed VSLAM system can robustly estimate the robot’s pose in complex real-world scenarios,construct point cloud map with rich geometric information,and improve positioning accuracy of the system.