Research On Visual-inertial Localization With Two-dimensional Code Markers

For unmanned driving,it is very important to realize high-precision localization in real time,and the most common method is localizing by satellite.However,when the signal of satellite is weak or even lost,achieving high-precision localization is difficult.By virtue of low cost and easy integration,visual odometry which based on camera has been widely studied.However,visual odometry is very sensitive to illumination conditions and difficult to adapt to complex scenes.Therefore,the inertial measurement unit can be added to solve this problem.Visual-inertial localization method will produce cumulative errors,so it is not suitable to be used in large-scale experiment.In order to eliminate cumulative errors and ensure the accuracy of localization,this paper adds two-dimensional code markers to assist localizing based on visual-inertial localization method.Firstly,building visual-inertial odometry front-end.As for visual measurements,Harris corner points in the image are extracted,then using LK sparse optical flow algorithm to track features.The random sampling consistency algorithm is used to remove mismatches.After that,the camera pose is estimated according to the matched features.As for inertia measurements,the pre-integrated processing is carried out,the pre-integrated model is established,and the formula is deduced.The method of visual-inertial initialization and alignment are studied.Then,in order to remove the noise of the odometry and improve the accuracy of localization,nonlinear optimization algorithm is carried out at the back-end.In order to reduce the computation and keep the system running in real time,only the key frames are calculated,and the state of key frames is managed by sliding window.The algorithm of selecting and marginalizing key frames is studied.Establishing visual residual model and inertial residual model,and the back-end objective function is established based on the nonlinear optimization algorithm.Then,in order to eliminate the cumulative errors caused by visual-inertial localization method,adding two-dimensional code markers to assist localizing.In order to reduce the calculation and remove irrelevant features and noise in the image,preprocessing the image.The relevant algorithms about preprocessing are studied and compared.Then locating the area of two-dimensional code in the image and decoding the code.Finally,estimating the camera pose according to the position information of two-dimensional code.Finally,experiments are designed to verify and analyze the localization algorithm.As for visual-inertial localization method,both open source data set and real-world scene are used.The experimental result shows that visual-inertial localization method will produce cumulative errors.The real-world scene experiment is carried out on the visual inertial localization method which integrates two-dimensional code marker positioning.The experimental result shows that the accuracy of the localization system is improved when integrating two-dimensional code markers,and cumulative error is eliminated.