Visual-inertial Odometry Using Point And Line Features

Visual inertial odometry(VIO)is widely used in unmanned aerial vehicle,autonomous driving,VR/AR and other fields to solve the problem of real-time localization when mobile terminals runs in the new environment,and has important research significance.In order to improve the accuracy and robustness of pose estimation,this paper proposes a tightly-coupled visual inertial odometry with point and line features using optimization method.In man-made environments,there are a lot of line features.Adding the observation of the line features in the environment to the VIO system can provide enough visual information constraints between the camera frame at low-texture scenes.The main research contents and research methods of this article include:1.In the point feature tracking module,to solve the homography matrix and illumination change parameters between the continuous image frame,this paper minimizes the photometric error of all image patches with the optimization-based method.Then,use the homography matrix to predict the position of the feature point in the current image frame.Finally,KLT optical flow method be used to track the precise position of the feature point.This method can obtain a higher tracking success rate.In the line tracking process,this paper adopts the method of tracking the sampling points of the straight line first,and then find the right results.2.A new IMU pre-integration method is used in this paper.The pre-integration state of the IMU is first integrated by the median integration method,then a discrete form of error state equation is derived from the integration result.The median value of the IMU pre-integration and the covariance of the error state are obtained by the error state equation.The reliability of this pre-integration method is confirmed from the final pose estimation results in this paper.3.In the sliding window optimizer of VIO,to estimate the system states,this paper uses the method of nonlinear optimization to minimize visual point feature observation residuals,visual line feature observation residuals,IMU pre-integration observation residuals and a priori information residuals.In order to keep the constant number of keyframes in the sliding window,this paper presents the selection strategy of marginalized keyframes and the selection method of landmark points and lines related to marginalized keyframes.4.In the loopclosure detection and pose graph optimization module,this paper uses DBo W2 to detect loop candidate keyframes.This paper gives a method for matching feature points between the images of the current keyframe and the keyframe of the loop keyframe local map.Then,the relative pose is found between the current keyframe and the local map of the loop keyframe.Finally,the drift of poses is reduced through the pose graph optimization.