Research On Monocular Visual SLAM/INS Combined Localization Method

With the continuous development of computer vision and navigation technology,visual SLAM technology with camera as the main localization sensor has become the mainstream research direction.In the complex indoor and outdoor environment,a single vision sensor is difficult to ensure long-term stable work,so the positioning method of data fusion between the camera and other sensors has attracted many scholars to study,and combined positioning has been more and more applied.The inertial navigation system is complementary to the monocular camera due to its fast response and estimable absolute scale.Based on the analysis of the localization with monocular camera,this thesis studies the monocular visual SLAM/INS combined localization method.The main work in this thesis can be summarized as follows:(1)The basic theory of monocular visual SLAM system is introduced,including common coordinate system and its transformation relationship,image feature extraction and tracking matching,motion recovery of monocular camera,loop detection,etc.This thesis introduces the basic theory of INS positioning,including common coordinate system and its transformation relationship,state solution and error transfer.(2)The calibration theory of monocular camera,inertial navigation and the joint calibration theory of monocular camera-inertial navigation are introduced.The calibration experiments of monocular camera,inertial navigation and joint calibration of monocular camera-inertial navigation are carried out,and the experimental results are analyzed.It provides the necessary initial parameters and precision guarantee for the monocular visual SLAM/INS integrated positioning system.(3)Aiming at the problem that the absolute scale cannot be obtained by monocular vision,a scale factor estimation method based on least squares was proposed.The scale factor of each frame was solved by using the same speed of two sensors at the same time,and the optimal scale factor was obtained by taking multiple groups of data and using the least squares method.This method can estimate the absolute scale of monocular vision effectively.(4)Aiming at the problem that the number of images affects the accuracy and real-time performance of the integrated positioning system,a keyframe selection strategy of monocular vision is proposed.Different time and distance thresholds are adopted for different experimental scenes,which can reduce the calculation amount of back-end optimization and improve the positioning performance of the integrated system.(5)The integrated localization theory of monocular visual SLAM/INS was studied,and the tight-coupling optimization model was established.The modules of IMU pre-integration and error transfer,back-end optimization based on sliding window,closed-loop detection and global optimization were deduced.(6)The keyframe selection strategy and the monocular visual SLAM/INS combined localization method in this thesis are verified experimentally.The combined localization experiments are carried out in the indoor small scene and the outdoor large scene respectively,and the experimental results are compared and analyzed.The experimental results show that the proposed key frame selection strategy is feasible,and the designed combined positioning method is suitable for different scenes.When the feature tracking fails,the continuity of positioning can be guaranteed,the cumulative error of a single sensor can be effectively reduced,and the positioning accuracy and robustness are better.The thesis has 54 figures,7 tables and 81 references.