Research Of Visual SLAM Based On IMU Fusion

Visual-SLAM plays a vital role in autonomous navigation and localization for Robots.Compared to Stereo Vision and RGB-D SLAM,the advantage of mono-SLAM lies in the adaptiveness to the environment with different scales.However,it is a challenge for Mono-SLAM to restore the absolute scale of the environment;thus,a sensor fusion method should be applied to address this problem.This article restores the absolute scale missed in Visual-SLAM by fusing IMU and enhances its robustness.Considering different error models,SLAM front-end is categorized into Direct Method(brightness error)and Feature-based Method(reprojection error).This article realizes the feature matching and closed-loop detection based on the feature points and accelerates the matching procedure by using dictionary;then,the epipolar geometry theory is introduced to estimate the relative pose between the adjacent cameras,which is also borrowed to monitor the integration procedure of the IMU between frames.Besides,if the front-end fails to estimate the pose,by solving the PnP problem,relocalization can be conducted for restoring the tracking of the pose.In the state-of-art IMU and visualized SLAM fusion algorithms,state parameters such as position,velocity,zero offset and posture etc.are usually chosen as the optimization parameters;however,outcomes in such models are prone to be divergence.This article proposes a method optimizing over only state parameters of pose,velocity and accelerator zero offset.The pose is obtained by IMU AHRS,which reduces the initial camera calibration procedure before fusing IMU and turns out that this method has a strong robustness.This article builds a corrective model with high accuracy using the noise model of gyroscope,accelerator and compassor.The Error State Kalman Filter Model is built over the data collected from the sensors,resulting in high accurate AHRS results.After obtaining the AHRS results of camera postures,the measure results collected from IMU accelerator and cameras are combined to build an optimization model with respect to pose,velocity and the accelerator zero offset,and the model is optimized within an optimization window.A sliding window is adopted to replace the earlier state variable,marginally,with new ones.A sliding window optimization method is used to fuse the local measure results of cameras with IMU,and there is still a drift between the moving trajectory and the map.During closed-loop detection,a dictionary accelerating method is used to obtain the closed-loop matching between frames and the bundle adjustment optimization model is used to execute a global optimization.The experiment results illustrate the effectiveness of the closed-loop detection and optimization in addressing the drifting problem.