Research On Positioning Method Based On SINS/monocular Vision/Lidar Combination

To adapt to the complex terrain of the influence of the electromagnetic environment,it is necessary to use the carrier itself to carry sensors to achieve passive autonomous positioning.Regardless of the inertial navigation system,visual sensor,or lidar,they all have their advantages and disadvantages,and they all have certain limitations in the application.This paper focuses on solving the positioning problem of light uncrewed vehicles in the battlefield environment.By leveraging the strengths of each sensor,it makes up for the lack of accuracy of a single low-cost sensor,thereby improving the accuracy and robustness of the system and ultimately improving the applicability of the system in the battlefield environment.The major components of the study are as follows:(1)To address the problem of the performance degradation of the IMU due to the combined interference of the internal device noise and the external vibration noise when the carrier is moving,this paper studies the accelerometer calibration method and the gyroscope random error compensation method.By establishing a deterministic error model of the accelerometer for calibration,this paper will design a full-temperature and full-speed test.The test results show that the accelerometer error after calibration based on the established model is significantly reduced.Furthermore,based on the data-driven method,this paper proposes an algorithm that uses the Bi LSTM network to characterize its random error model and combines the EM algorithm for filter compensation.The random vibration test is designed to collect gyroscope data as the training set of the network.The gyroscope data in the proposed Bi LSTM-EM algorithm is used to verify the compensation effect.The results show that the proposed algorithm's maximum relative error after compensation is 3.56%,while the maximum relative errors between the original data and the ARMA model are 29.97% and 21.24%,respectively.Since the maximum relative error tends to occur in the low angular rate range,random errors have a more significant impact when the angular rate is low.It can be seen that the proposed Bi LSTM-EM algorithm has a certain effect on the gyroscope random error compensation method.(2)In the Le GO?LOAM algorithm,the effect of ground segmentation by setting a threshold in general and the constraint in the pose is insufficient.This paper proposes a ground segmentation method based on iterative plane fitting to establish a reasonable constraint.The results of the KITTI data set show that compared with the ALAOM algorithm and the Le GO?LOAM algorithm,the improved Le GO?LOAM algorithm proposed in this paper has better accuracy,indicating that the ground-based on iterative plane fitting and segmentation has better in the pose.Therefore,the odometer positioning effect is improved.(3)Aiming at the problem of mesoscale uncertainty in monocular visual odometry,and improved supervised depth prediction algorithm based on Unet++ and a monocular visual odometry algorithm based on fusion depth prediction are proposed.The proposed monocular depth prediction algorithm Res-Unet++ network uses Res Net101 as the encoder and designs the encoder-decoder structure to provide accurate depth values.The test results on the KITTI Eigen split dataset show that the proposed Res-Unet++ network outperforms the original Unet,Unet++,and Unet3+networks.Then the proposed monocular depth prediction algorithm is integrated into the SVO algorithm to provide a reliable initial value for its depth filter,reduce the uncertainty of the initial depth value in the original algorithm,and accelerate the convergence of the depth filter.The test results on the KITTI dataset show that,compared with the DSO algorithm and the ORB-SLAM algorithm with closed-loop detection,the Res-Unet++-SVO algorithm proposed in this paper is more effective for accurate trajectory tracking,demonstrated that the monocular visual odometry algorithm with depth prediction proposed in this paper can effectively solve the scale uncertainty problem of traditional monocular odometry.(4)We are aiming at the problem that the sensor failure rate increases in the battlefield environment.This paper adopts the federated Kalman filter for multi-sensor combined positioning.We simulate sensor failure by overlaying point clouds and images from a single frame in the KITTI dataset over a while after that.The results show that the system trajectory can still converge after a single sensor fails,indicating that the system has certain fault tolerance.