Research On Multi-sensor Fusion Positioning Strategy For Intelligent Vehicles Based On Graph Optimization

With the development of the economy and the continuous progress of science and technology,intelligent vehicles have become one of the most important research directions in the field of vehicle engineering.Positioning technology of intelligent vehicles is an important part of environmental perception and is also the basis for intelligent vehicles to realize decision planning and motion control.The breakthrough of high-precision positioning technology for intelligent vehicles is of great significance to the long-term development of the automotive industry.Presently,various independent positioning methods have their inevitable defects and poor environmental adaptability.Therefore,scholars at home and abroad have studied multisensor fusion positioning technology to obtain stable and reliable positioning information of intelligent vehicles.The hardware cost of the camera and inertial measurement unit（IMU）is low and their positioning characteristics are obviously complementary.Visual-inertial odometry（VIO）formed by the fusion of camera and IMU can achieve self-positioning in unknown environments without relying on external signal sources,but its positioning effect is easily affected by factors such as illumination,and the positioning robustness in large-scale outdoor environments is poor.In addition,both camera and IMU positioning belong to relative positioning,and their positioning errors will continue to accumulate,which cannot meet the positioning requirements of intelligent vehicles when they run for a long time.It is necessary to rely on absolute positioning technology to eliminate the accumulated errors of VIO.The global positioning system（GPS）widely equipped in intelligent vehicles is an absolute positioning technology,which can provide real-time,all-weather,and full-time positioning services.However,its positioning accuracy can only reach the meter level,which cannot meet the accuracy requirements of intelligent vehicles.Therefore,GPS usually needs to be equipped with low-cost network real-time kinematic（NRTK）positioning technology to improve accuracy.However,GPS/NRTK will have very low accuracy in occluded areas and may even lose positioning,which cannot guarantee the stability and continuity of positioning.This means that GPS/NRTK needs to be fused with self-positioning systems that do not rely on external signal sources.Therefore,this paper proposed a multi-sensor fusion positioning strategy（Camera/IMU/GPS/NRTK）based on graph optimization method to improve the positioning ability of intelligent vehicles in large-scale unknown outdoor environments and conducts KITTI dataset simulation and campus real vehicle test.Overall,the main contributions of this study are as follows:（1）Front-end Processing and Initialization Algorithm of Visual Inertial OdometerThe front-end processing and initialization algorithm of visual-inertial odometry was established.The principle of camera imaging was analyzed and modeled,and the image feature extraction algorithm based on the FAST corner point,the feature tracking algorithm based on the image pyramid,and the feature matching algorithm based on the RANSAC algorithm were established.The frame adaptive selection strategy used the KITTI data set and the data collected from the real vehicle to verify the image feature processing algorithm.The IMU pre-integration model was derived based on the established IMU model to reduce the amount of calculation.This paper performed depth estimation on 3D landmarks and estimated camera pose based on triangulation and Pn P.Based on camera measurements and IMU measurements,an objective function was constructed to achieve visual and inertial alignment,and parameters such as scale,gyroscope bias,initial velocity,and gravity vector were estimated.（2）Backend Algorithm of Visual Inertial Odometry Based on Graph OptimizationA back-end algorithm for visual-inertial odometry based on graph optimization was established.The sliding window method based on graph optimization realized the tight coupling of the camera and the IMU,and at the same time derived the measurement residual of vision and IMU,the Jacobian matrix of the measurement residual,and the covariance matrix of the incremental error.The marginalization strategy based on Schur’s theory was established to limit the scale of BA optimization,thereby improving the computational efficiency of the whole system.A loop closure detection optimization algorithm based on the bag-of-words model was established.The loopback frame was judged by similarity,consistency,continuity,geometry,etc.and the loopback information was added to the overall BA optimization function for loopback detection optimization,which could effectively eliminate accumulated errors and improve positioning accuracy.（3）Multi-sensor Fusion Positioning Strategy Based on Global Pose Graph OptimizationThis paper proposed a multi-sensor fusion positioning strategy based on global pose graph optimization.The spatial coordinate transformation equation and IMU pre-integration method were used to realize the time synchronization and space synchronization of different sensors.Secondly,the architecture of the multi-sensor fusion positioning strategy was established,which mainly included GPS/NRTK and VIO.At the same time,the NRTK positioning and error correction principle were analyzed.This paper established an adaptive fusion algorithm for VIO and GPS/NRTK based on the global pose graph optimization theory,which could effectively improve the positioning accuracy of the entire system and could meet the positioning requirement of autonomous vehicles.（4）Experiment and AnalysisAccording to the test requirements,the KITTI simulation platform and the outdoor positioning real vehicle platform were established and the algorithm evaluation indicators were selected.The KITTI data set was preprocessed and used to verify the algorithm in a largescale outdoor environment simulation test.The relevant experimental equipment in the real vehicle platform was calibrated and the accuracy of the key equipment was evaluated.Based on the real vehicle platform,the algorithm proposed in this paper was verified and analyzed.The test results showed that the multi-sensor fusion positioning strategy based on graph optimization proposed in this paper had the advantages of high performance,low cost,etc.,and could meet the requirements of intelligent vehicles.