Research On Monocular Visual SLAM Technology Based On Multi-source Information Fusion

Simultaneous Localization and Mapping(SLAM)is a key technology for intelligent mobile robots and is widely applied in fields such as autonomous driving,virtual reality,and environmental mapping.Visual SLAM that using a monocular camera has advantages such as small size,lightweight,and low cost,which has attracted increasing attention.However,in scenes with textureless regions,rapidly changing lighting,and intense motion,the robustness of monocular visual SLAM is poor,and it cannot recover scale information,limiting its application range.This thesis mainly studies how to improve the positioning accuracy and robustness of monocular visual SLAM in complex scenes through multi-source information fusion,including multimodal sensor information,environmental geometric features,and carrier kinematics information.The main research content is as follows:(1)This thesis proposes a lightweight visual frontend algorithm based on IMU prior information to improve the system’s feature tracking ability in dynamic environments.First,the image is iridized to extract corners,and then the prior pose obtained from IMU calculation is used to assist feature tracking,outlier rejection,and map point triangulation to obtain more accurate visual geometric constraints.The experimental results show that the proposed algorithm obtains more robust feature association and higher pose estimation accuracy in high-dynamic scenes.(2)To solve the degradation problem of monocular visual-inertial SLAM on wheeled robots,this thesis proposes a tightly coupled visual-inertial-wheel-speed SLAM algorithm based on factor graph optimization,which can provide robust and high-precision pose estimation results for ground wheeled robots.Firstly,based on the vehicle’s kinematic information and the original observation of the wheel speed sensor,a manifold-based IMU/ODO joint preintegration model is proposed to solve the problem of increased computational cost caused by repeated integration.Then,the residuals and Jacobian matrix of the preintegration are derived,and the intrinsic parameters of the wheel speed sensor are calibrated online in the backend optimization.Experimental results in a large-scale urban car environment show that the proposed algorithm has higher positioning accuracy and robustness compared to mainstream visual-inertial SLAM methods.(3)To improve the positioning accuracy of point-feature-based visual SLAM algorithms in scenes with fast motion,weak texture,and rapidly changing lighting,this thesis introduces line features as a supplement.First,an improved ELSED line detection algorithm and LBD descriptor are used to extract and match line features in the image.Then,the Plücker coordinates and orthogonal representation method are used to parameterize and calculate the geometry of the lines to solve the problem of unstable calculation results caused by over-parameterization.The residual model and Jacobian matrix of spatial lines are derived,and a sliding window strategy is used to perform nonlinear optimization on point features,line features,and IMU preintegration residuals.The experimental results show that compared with methods relying solely on point features,the proposed algorithm can obtain more accurate pose estimation results and construct a structured map of the surrounding environment.