Towards Vision Aided Inertial Navigation Techniques For Ground Vehicles

Manned or unmanned ground vehicles with autonomous ability have attracted people's attention greatly in recent decades.As a result,there is an increased demand for the navigation performance improvement of low cost navigation systems.The integration of INS and GNSS receivers is well-known and commonly used in ground vehicle applications.However,GNSS can suffer from obstruction and multipath effect errors in city canyons,tunnels,woodlands and mountainous regions,and is also prone to the possibility of being jammed or spoofed.Therefore,the navigation in GNSS-denied environment is of great interest among a lot of researchers.It is significant to study the methods to mitigate the error drift by using low cost navigation sensors and aiding sensors,as well as the new integration schemes and techniques,especially using knowledges from multiple disciplines.It has long been one of the hot topics in the navigation area.The content of this thesis is as follows.1.In the non-holonomic constraints(NHC)and odometer(OD)aided navigation system,the boresight error and lever-arm of IMU with respect to the vehicle frame are fully considered in the system model.Considering the characteristic of low cost IMU sensors,the observability of INS/NHC/OD integration is theoretically analyzed,which is different from the existing analysis using an inertial-grade INS.To obtain higher IMU-vehicle calibration accuracy,the fusion scheme based on the Unscented Kalman filter(UKF)is developed,taking special treatment on the unscented transform to the quaternion.Simulation test shows that UKF outperforms EKF in estimating the calibration parameters,especially when the boresight error is slightly larger.2.The vanishing point(VP)aided INS methods are thoroughly studied based on the parallel lane marking observations from a forward looking camera.First,the relationship between vanishing point coordinates and relative attitude of the camera with respect to the road is mathematically developed.Based on this,the relative heading formula is derived.The whole VP aiding scheme is proposed,including straight lane detection,uncertainty analysis,sequential Kalman filtering,and sensitivity analysis of INS/VP integration.The AIME(Autonomous Integrity Monitored Extrapolation)soft failure scheme is adopted to detect the small curve of the lane.The algorithm is tested by simulations and experiments.It is shown that with additional help of VP,33% improvement of the positioning accuracy is achieved than INS/NHC alone,reaching 0.32% DT(distance traveled).3.The relative pose calculated by 2D-2D visual odometry(VO)of a monocular camera is utilized to aid the INS.The frame to frame relative pose is calculated based on the epipolar constraint.An uncertainty estimation method for the relative attitude from the vision system is developed,which is essential for the sensor fusion.Simulations and experiments show the validness of the covariance estimation method.A simple but effective failure detection of the VO system is proposed based on the translation vector from VO.Finally,the loosely coupled INS/NHC/VO integration is developed,and the observability analysis proves the complementary properties of INS/NHC and INS/VO integration.The experiments show that in the INS/NHC/VO integrated navigation,the average horizontal positioning RMS error of 4 experiments is within 0.30% DT.4.The line features observed by a camera are extracted and parameterized for further improving the accuracy of existing VINS.The first approach is to extract the lines corresponding to the vertical 3D lines of buildings and thus to calculate the roll angle of the vehicle.This helps the existing point feature based VINS using Multiple State Constraints Kalman filter(MSCKF).Furthermore,a new straight line parameterization,which is called AIDPL,is proposed for the undelayed initialization of 3D space lines when using line based VINS under the framework of EKF-SLAM.The Monte Carlo simulation tests demonstrate that the positioning accuracy is significantly improved using proposed tightly-coupled VINS.Meanwhile,the 3D lines in the environment are estimated effectively and quickly in the setup.