Research On Stereo Visual Inertial SLAM Algorithm Based On Filtering And Nonlinear Optimal Combination

With the development of computer vision,inertial navigation and multi-sensor fusion technology,the use of visual inertia simultaneous positioning and map construction Simultaneous Localization and Mapping algorithm to solve the problem of unstructured scenes,GPS signal weak observation conditions,the robot The problem of localization and pose estimation has gradually become a hot research direction in the SLAM field.The performance of visual inertial SLAM will be affected by scene texture characteristics,lighting changes,carrier movement speed and sensor noise.In order to reduce the influence of the above factors on the result of the pose measurement,the SLAM algorithm usually adopts filtering or nonlinear optimization methods to reduce the accumulated error.However,methods based on filtering have a large amount of linearization errors,and methods based on nonlinear optimization will occupy a lot of computing resources,and it is difficult to ensure real-time performance on small mobile platforms.Considering these issues,this paper proposes a binocular visual-inertial SLAM algorithm that uses filtering and nonlinear optimization at the same time.The main research work of the thesis includes:1)The method of visual feature extraction and matching of visual inertial SLAM is studied,and an algorithm for visual feature extraction through BRISK method is proposed,and then feature tracking algorithm is used for LK optical flow method improved by gradient gold tower.Aiming at the problem of repeated integration of IMU caused by the different sampling frequency of IMU and binocular camera,the derivation process of pre-integration of IMU is introduced,and the premise of tight coupling of visual inertial measurement data is provided.2)A close-coupled binocular visual-inertial SLAM algorithm using both extended Kalman filter(EKF)and Incremental Bundle Adjustment(IBA)is proposed.At the front end,EKF is used to tightly couple the visual inertial measurement information to establish a multi-state constraint model,and information augmentation is added to the traditional EKF update,which effectively improves the estimation accuracy.In order to reduce the linearization error caused by EKF,in the back-end design,the incremental beam-planning method with global BA and local BA with sliding window is used to optimize the pose estimation result,and the key frame and the key frame optimized by the front-end EKF are used.Marginalization reduces the occupation of computing resources.And design loop detection to ensure the robustness of pose estimation in long-distance loop motion.3)A mobile robot platform with ZED binocular camera and IMU as sensors was independently built to verify the algorithm.Compared with the SLAM algorithm that only uses the filtering method or the optimization method on the public data set.The experimental results show that the algorithm in this paper can reduce the linearization error of the pose estimation,improve the estimation accuracy,and reduce the occupation of computing resources.A mobile robot is used to perform pose estimation in real-time in an indoor environment of about 20 square meters.The results show that the practical feasibility of the proposed algorithm is verified.