Research On Tightly-coupled Monocular Visual-inertial Odometry With Vehicle Motion Constraints

For intelligent and connected vehicles,the ability to obtain accurate vehicle pose in real time and stably is a necessary condition for autonomous navigation and autonomous movement.The pose estimation technology based on vision sensor has the advantages of low cost and convenient hardware layout,so it has attracted extensive attention in the research of high-precision positioning of intelligent and connected vehicles.However,the monocular camera cannot recover the scale information of motion and environment,and when encountering areas lacking texture or dynamic environments,the low-level feature-based method is used for its visual information processing,resulting in poor system robustness or even failure.Based on the pure visual pose estimation method,the visual-inertial fusion pose estimation method integrates the acceleration and angular velocity measurement information of consumer-grade inertial sensors,and has better accuracy and robustness on the basis of maintaining the cost advantage,and has achieved a lot of research progress and practical applications in the field of pose estimation of mobile robots and autonomous vehicles.In the visual-inertial fusion positioning system,the visual observation and inertial measurement are tightly coupled to estimate the 6-DOF pose of the carrier,which can also produce better estimation results when the carrier moves rapidly or in the case of severe illumination changes.When applied to ground vehicles,the traditional monocular visual inertial pose estimation method cannot observe the degenerate motion under the conditions of uniform linear motion or uniform circular motion,and the observability changes.In addition,due to the limitation of the system nonlinearity and the inability to directly observe the distance,the initialization process of the monocular visual inertial pose estimation method under the lack of excitation of the inertial measurement element has always been a huge challenge: on the one hand,monocular vision cannot estimate the metric scale of the environment and its own motion due to its own measurement limitations,and pure visual pose estimation algorithms are less robust in fast-moving or dynamic environments;on the other hand,during initialization,insufficient excitation of the inertial measurement unit will cause the gyroscope and accelerometer bias estimates to generally struggle to converge,which in turn leads to the failure of the initialization method based on the loose coupling of visual-inertial measurement.In this case,the initial state of the whole system,i.e.scale information,carrier velocity,gravity vector and bias cannot be estimated quickly and accurately.Compared with the limitation of scale observation by visual and inertial sensors under special working conditions,vehicle odometry can stably obtain vehicle motion measurement with intuitive scale information.How to effectively fuse consumer-grade visual inertial sensors with vehicle odometry that can accurately reflect vehicle motion,so as to improve the accuracy and robustness of the pose estimation of the multi-sensor fusion positioning system under various working conditions,is a technical research problem that is very valuable and has broad application prospects.In order to solve the problems of poor robustness,low accuracy and possible initialization failure of traditional pose estimation methods based on visual-inertial fusion when applied to ground vehicles,this paper proposes a tightly coupled monocular visual-inertial odometry with vehicle motion constraints.This method uses the high-frequency vehicle motion information obtained from the vehicle control unit by the CAN bus,uses the vehicle kinematics model to estimate the linear and angular velocity states of the vehicle,integrates the measured values,and uses a sliding window-based nonlinear optimization method to process tightly coupled monocular vision,inertial measurement unit and vehicle odometer measurement constraints and estimate vehicle state.Tightly coupled measurement constraints help to estimate accurate vehicle motion states,and the introduction of vehicle motion information makes the system more robust for autonomous driving applications.The main contribution of this paper is to propose a multi-sensor fusion localization framework suitable for intelligent and connected vehicles that tightly couples monocular vision,inertial measurement unit and vehicle odometry based on vehicle kinematics model,and proposes a robust system initialization method,which can bootstrap a slidingwindow-based nonlinear optimization solver from unknown initial states.In addition,in view of the systematic error of the vehicle odometry measurement process based on the vehicle kinematic model,a method for estimating the system error parameters of the vehicle motion model based on the error state kalman filter is proposed.The internal and external parameters of the vehicle odometry are calibrated,and the accuracy of the algorithm is evaluated through real vehicle experiments under various working conditions,which proves the effectiveness of the system error parameters and the external parameter estimation algorithm.Aiming at the multi-sensor fusion positioning algorithm proposed in this paper,a comparison experiment of various algorithms is carried out in the small scene environment of the park and the open large scene environment based on real vehicles.The experiments show that,compared with the existing visual inertial navigation methods,this paper proposes the algorithm achieves higher localization accuracy and robustness.