Research On Indoor Localization Technology Based On Visual Point-Line Features And IMU Fusion

Indoor localization technology has become one of the popular research directions in the field of mobile robots.The key to this technology is the accurate estimation of the robot’s motion state in real-time in an unknown environment.Visual-Inertial Odometry(VIO)is one of the methods to solve this problem and has been widely studied and applied.This paper addresses the issue of reduced localization accuracy due to the failure of point feature extraction and tracking in low-textured,and low-light environments,which are based solely on the fusion of visual point features and inertial navigation.By introducing the method of incorporating indoor environment line features,we utilize visual point-line feature processing algorithms to extract and track pointline features and construct an objective function containing point-line-IMU residual information.By optimizing this objective function,we aim to improve the localization accuracy and robustness of the system.The main research work of this paper includes:(1)To address the issue of mismatched point features during the matching process,an improved binary scale-invariant feature algorithm(BRISK)is proposed.To tackle the problem of time-consuming line feature detection affecting the real-time performance of the VIO system,a novel fast line feature processing algorithm—EDlines is adopted.The improved BRISK algorithm refines the scale space by using continuously varying scale parameters to obtain different scale image information,making the algorithm more adaptable to different image scales and rotation angles,and enhancing the accuracy,robustness,and reliability of the algorithm.The robustness of the improved BRISK algorithm and the solution to the mismatch problem in the feature point matching process are demonstrated through dataset experiments and point feature extraction and matching tests.The EDlines line segment extraction algorithm significantly improves the speed of line segment detection and matching while maintaining the accuracy of line segment detection and matching.By comparing the performance of EDlines and LSD algorithms on the EuRoC dataset,it is shown that the EDlines algorithm is at least 6 times faster than LSD while having a comparable line segment detection and matching performance.(2)To address the issue of accurately estimating the robot’s pose,a pose estimation method based on a visual point-line and IMU fusion optimization model is proposed.First,the representation of lines in three-dimensional space is introduced,and point feature residual models,line feature residual models,and IMU residual models are constructed.Second,as the system continuously explores the environment,more and more variable information is added to the system.As the dimensions increase,the computational complexity gradually increases.To reduce the computational complexity of the system,a sliding window-based marginalization strategy is adopted.The marginalized visual information and inertial information from the images are transformed into prior information and combined with IMU measurement errors and point-line reprojection errors to construct an objective function.The mobile robot’s optimal pose estimation is achieved through iterative nonlinear optimization of this objective function.By conducting simulation experiments on the EuRoC dataset sequence,the localization scheme presented in this paper is compared with PL-VINS and VINS_mono systems in terms of localization accuracy and real-time performance.A comprehensive analysis of algorithm runtime,absolute trajectory root mean square error,and other parameters demonstrates that the proposed localization scheme has high localization accuracy and real-time performance.(3)An experimental platform based on the EAIBOT mobile robot is set up,and the overall framework of the localization system is introduced,mainly including data preprocessing,system initialization,sliding window nonlinear optimization and pose estimation,and loop closure detection.To verify the localization accuracy and feasibility of the proposed localization scheme in a real environment,the localization scheme is compared with the PL-VINS system by conducting localization experiments on the experimental platform following indoor predefined trajectories.The experimental results show that the localization trajectory position error mean values for the proposed localization scheme under square and rectangular routes are 1.34%and 1.23%,respectively,which are better than the PL-VINS system’s position error mean values of 1.87%and 1.98%.The comprehensive comparison demonstrates the accuracy and feasibility of the proposed localization scheme.