Research On Visual-Inertial SLAM Based On Point And Line Features Of Mobile Robot

Mobile robot can work in execrable environment and complete dangerous works instead of people,which makes new demands on its automation and intelligence.In order to achieve highprecision robot localizations,a simultaneous localization and mapping(SLAM)system that adds point and line features to the visual-inertial odometer is proposed.A complete slam framework is built through multi-sensor fusion and strong constraints of line features.Then an experimental platform is built to verify the accuracy and real-time performance of the system.As a result,the high-precision real-time positioning of the robot in indoor and corridor environment is realized.In this paper,the joint initialization of monocular camera and IMU is carried out,and the camera imaging model and IMU pre-integration model are established respectively.The information of camera and IMU is fused through the tight coupling of sliding window,and the multi-sensor fusion is realized,which provides an accurate initial value for the feature tracking of front-end visual odometer.By introducing additional line feature constraints,a method to improve the accuracy of slam system is proposed.Firstly,the segment extraction algorithm is improved,through the scheme of segment extraction and matching using the improved LSD+LBD algorithm is determined.Secondly,the rotation and reprojection transformation of line features are parameterized by pruck coordinates at the front end,and the orthogonal form of pruck coordinates is used for parameter optimization at the back end.Deriving the expression of line feature error and its Hessian matrix and information matrix.Finally,the comprehensive point line optimization factor graph is established according to the sliding window,and the additional constraints are obtained.The point and line feature is introduced into the visual-inertial odometer,VI-BLSD-Bow(Visual-IMU-BRIEF-LSD-Bag)framework which can realize autonomous navigation in unknown environment is proposed.Firstly,the hidden parameters of LSD algorithm are improved to realize the real-time extraction of line features.Secondly,a different matching and screening framework of point line features is proposed to reduce the generation of false matching.Finally,a loop detection module is added to realize the calibration of cumulative error.The upper computer of the robot car is connected with the camera and IMU respectively to form the experimental platform,so as to form an experimental platform.Firstly,the joint calibration experiment of camera and IMU is carried out to eliminate the influence of camera distortion and IMU bias on the measurement accuracy of the sensor;Secondly,the visual word bag dictionary model is retrained,so that the system can calibrate the cumulative error and improve the global consistency of the mapping trajectory.Thirdly,the recognition performance of word bag structure to loop detection is tested.Finally,experiments are carried out in indoor and outdoor,showing that the accuracy and robustness of VI-VLSD-BOW system are better than VINS-Mono.