Research On Binocular-Vision Simultaneous Localization And Mapping Algorithm Based On Multi-Sensor Fusion

The ability of accurate positioning and perception of the environment are the premise of intelligent mobile robot to work normally.With the deepening of research,the technology of Simultaneous Localization and Mapping(SLAM)came into being.Using a single sensor to locate the robot is difficult to ensure its long-term stable work in various environments.More and more multi-sensor fusion localization solutions are being researched and applied.The visual sensor is prone to lack of feature points in short-term rapid motion or in non-textured area,which cannot guarantee good performance of feature tracking and pose estimation.Inertial Measurement Unit can make up for this deficiency,and the visual sensor information can also correct the influence of IMU drift.Visual Inertial Odometry integrates the data of camera and IMU sensors.With its high accuracy,high complementarity and high robustness,it has become an important research topic in slam field.However,the amount of visual information is so large that the amount of calculation after the data fusion of multiple sensors is also large,making the research of VIO still have many challenges.This thesis introduces the theories of camera distortion correction model.IMU error model and the transformation between various coordinate systems.The camera and IMU are calibrated jointly,laying the foundation for the later experiments.The point features in visual SLAM have been researched and applied more maturely.However,there are a large number of line features in the structured environment.which can be used as a supplement when point features are missing or insufficient.Therefore,this paper proposes a visual odometry combining point features and line features,studies the point feature extraction and matching algorithm,the line feature extraction algorithm and the parameterized model.The advantages of point features and line features complement each other to improve the robustness of the system.In order to solve the problem of repeated integration of IMU in the process of pose solving.IMU pre-integration is used to reduce the computational complexity.The nonlinear optimization method is used to minimize IMU pre-integration observation residuals,point feature observation residuals,line feature observation residuals and prior information residuals to estimate the state of the system.In order to maintain a constant number of key frames in the optimized sliding window,the key frame selection strategy and marginalization strategy are introduced.A hybrid dictionary tree construction method based on the combination of point features and line features is proposed for closed-loop detection.Finally,the overall framework of visual inertial SLAM based on the combination of point features and line features is established.