| Providing real-time and accurate vehicle pose information is a necessary prerequisite for intelligent and connected vehicle navigation.Visual Inertial Odometry(VIO)possesses the advantages of favorable robustness,high concealment,miniaturization,and low cost.It has been widely used in the pose estimation of unmanned aerial vehicles(UAVs)and mobile robots.However,there are still lots of limitations in applying VIO to the ground vehicles.For example,the vehicle is often under constant acceleration and straight driving conditions or approximately combined two driving conditions.Under these specific conditions,the observability of the VIO system will change,resulting in the positioning accuracy of the vehicle is low even positioning fails.The urgent scientific problem is how to effectively integrate the three methods of pose estimation based on vehicle kinematics,vision,and inertia in the case of the Global Navigation Satellite System(GNSS)signal interference to improve the accuracy of vehicle pose estimation and the observability of the system.In view of the above problems,the researching focus in this paper is mainly on vehicle kinematic error estimation and extrinsic parameters calibration,formulation of VIO method based on vehicle kinematic error state model and VIO method based on vehicle kinematic error measurement model.The main contributions are described as follows:1)In order to address the vehicle positioning trajectory deviation caused by the systematic error of vehicle wheel odometry,a three-dimensional space vehicle kinematic error state model and a combined inertial and GNSS navigation system(GNSS-INS)position and angular velocity measurement model were proposed.Moreover,the observability of the proposed method was analytically derived.The experimental results show that the absolute trajectory error of wheel odometry with the vehicle kinematic error estimation is 86.60%,67.32%,56.53%,91.27%,79.10%,78.46%,69.83% and 61.68% lower than that of wheel odometry with ideal vehicle kinematic model on the eight real-world datasets,respectively.And the vehicle positioning trajectory deviation caused by the systematic error of the vehicle kinematics is effectively weakened.2)To address the inadequacy of the current spatial calibration accuracy of the vehicle coordinate system and the IMU coordinate system,a rough calibration of the rotating extrinsic parameters based on rotation sequence singular value decomposition and the GNSS-INS-assisted extrinsic parameter online calibration method were proposed considering the IMU bias.And the IMU error state kinematic model and GNSS-INS position measurement model were obtained as well as the analytically derived observability of the proposed method.The experimental results show that the absolute trajectory error of Stereo Multi-State Constraint Kalman Filter(S-MSCKF)with extrinsic parameter online calibration is 4.42%,12.25% and 2.69% lower than that of S-MSCKF calibrated by manual measurement on the three real-world datasets,respectively.Besides,the absolute trajectory position error of S-MSCKF calibrated by singular value decomposition and weighted quaternion method is 3.70%,9.41% and 1.53% lower than that of S-MSCKF calibrated by manual measurement on the three real-world datasets,respectively.3)Aiming at the limitation of pose estimation of vehicle kinematics model and the problem of VIO scale unobservable when the vehicle is under constant acceleration,a wheeled Visual Odometry(VO)system(WMO)was firstly proposed,where derived a vehicle kinematic error state model and the visual measurement model based on MultiState Constraint Kalman Filter(MSCKF).Then the gyroscope information was incorporated into the vehicle kinematic error state model,where a wheeled gyro VIO system(WGMO)was proposed.Subsequently,the observability of WMO and WGMO was analytically deduced,and the proposed algorithm was theoretically proved to be observable in the scale direction in the system.The experimental results show that the two proposed methods can solve the problem of scale unobservable of the vehicle under constant acceleration conditions with improved accuracy of the system’s pose estimation compared to the traditional S-MSCKF system.Compared with S-MSCKF on the seven real-world datasets,WGMO reduced the root mean square scale ratio of 27.79%,11.57%,30.60%、15.04%,20.71%,31.70% and 5.08%,WMO reduced the root mean square scale ratio of 32.52%,24.38%,37.30%,25.96%,26.10%,37.21% and 14.46%.Here need to note that the pose estimation accuracy and robustness of WGMO are better.Compared with S-MSCKF,the absolute trajectory position error of WGMO on the eight real-world datasets was reduced by 90.21%,4.11%,65.69%,87.70%,33.95%,87.66%,84.68% and 4.74%,respectively.Compared with OKVIS,the absolute trajectory position error of WGMO on the eight real-world datasets was reduced by 95.56%,79.17%,69.63%,92.52%,76.39%,88.48%,9.58% and 38.76%,respectively.4)To address the problem that VIO scale unobservable and the roll and pitch angles of the VIO method based on the vehicle kinematic error state model are not observable,this paper proposed a method based on vehicle relative kinematic error measurement model(ACK-MSCKF-M)and a monocular VIO(ACK-MSCKF-LM)method with consideration of the vehicle speed and yaw angular velocity errors taking into account the lever arm effect.Furthermore,the observability of ACK-MSCKF-M and ACKMSCKF-LM was analytically deduced.The experimental results show that the proposed two methods can improve the problem of scale unobservable of vehicles under the condition of constant acceleration while the roll and pitch angle of vehicles also can be observed.Moreover,ACK-MSCKF-LM overcomes the inconsistency of ACKMSCKF-M and further improves the accuracy of the pose estimation of the system.Compared with S-MSCKF,ACK-MSCKF-M reduced the root mean square scale ratio of 33.90%,18.57%,27.38%,36.76% and 18.54% on the five real-world datasets,respectively.And,ACK-MSCKF-M reduced the absolute trajectory position error of 86.29%,65.69%,74.84%,58.79% and 60.73% on the five real-world datasets,respectively.ACK-MSCKF-LM reduced the root mean square scale ratio of 28.70%,10.86%,19.48%,32.40% and 9.88% on the five real-world datasets,respectively.And,ACK-MSCKF-LM reduced the absolute trajectory position error of 91.55%,6.39%,71.97%,87.84%,33.33%,85.56%,79.69% and 64.29% on the eight real-world datasets,respectively.5)Based on the conclusion of observability analysis,the strategies with model order reduction of WGMO and ACK-MSCKF-LM were firstly proposed and the influence on positioning accuracy was studied.A comprehensive comparison and performance evaluation of the multi-sensor fusion positioning methods were conducted.The experimental results show that the strategies with model order reduction based on observable analysis have little influence on the vehicle positioning accuracy.Among these different methods,ACK-MSCKF-LM shows the best positioning accuracy on all of these real-world datasets.At the same time,compared with ACK-MSCKF-LM,WGMO saves accelerometer and steering wheel angle sensor,and has more advantages in terms of real-time performance and hardware cost.This paper can provide a theoretical basis and practical reference for the development and application of intelligent and connected vehicle positioning and navigation system. |
PDF Full Text Request