Research On EKF-SLAM Algorithm Using IMUs And Monocular Cameras

SLAM has been studied more than 20 years by the robotic community. Researchers hope to have the robot localize itself and draw a map of the surrounding somehow to realize the autonomously moving robot. On the other hand, a similar problem known as structure from motion (SFM) has been insightfully studied by the computer vision community for a long time, which is majorly about reconstructing 3D object form a bunch of images and estimating parameters of camera meanwhile. In the earlier years, vision has been introduced into studies to help addressing SLAM, which is termed as Vision-SLAM. Recently, EKF-based SLAM with the help of monocular camera and IMUs is intensely studied by robotic and computer vision community as the Micro-Electro-Mechanical Systems (MEMS) technology makes the sensors such as accelerometer and gyroscopes small and accurate enough, so that EKF-based SLAM using vision and IMUs can provide a better estimation of camera ego-motion and the position of landmarks in three-dimensional (3D) for all kinds of applications such as the navigation in GPS-denied situation and 3D reconstruction.In this master thesis, we wildly study the IMU-and vision-based SLAM problem, and focus on some stages of error-state KF-based SLAM system, including using quaternion to characterize rotation, the way the state and covariance matrix are propagated, the observation model, linearization of model, the parameterization methods in characterizing landmarks in 3D space, and the process of error-state KF. In particular, we address the issue of inverse depth parameterization and propose a re-parameterization technique to update the anchor point. From the experimental results, the proposed technique can improve the estimation of landmark position in some cases.