Research On Key Technology Of Low-cost GNSS/IMU/Camera Sensor Fusion For Navigation

With the continuous advancement of artificial intelligence and sensor technology,the application scenarios of navigation technology are expanding.More and more applications put forward higher requirements for navigation performance,especially the accuracy,reliability and availability of navigation parameters.At the same time,the volume,weight and cost of navigation systems are subject to more stringent restrictions.With the mature of integrated circuits and microelectro-mechanical technology,the performance of micro-sensors is continuously improved.Based on these sensors,the integrated navigation system can improve navigation performance,and reduce its size,weight and cost.Therefore,Low-cost GNSS/IMU/camera sensor fusion has become a domestic and foreign research hotspots.However,key technologies such as monocular visual-inertial SLAM and GNSS/IMU/camera fusion model for low-cost sensors still need to be further improved.Therefore,based on our low-cost GNSS/IMU/camera sensor fusion platform,this dissertation focuses on these two key technologies.The main work and innovations are as follows:1.As for the monocular visual-inertial SLAM technology,we improved the state-of-the-art monocular visual-inertial SLAM framework VINS-Mono in system initialization,sliding window optimization and loop optimization:(1)We suggested using the normalized vector displacement caused by camera translation additionally as the initialization success criterion,which improves the stability of the monocular visual odometry.(2)We proposed an approach to estimate the relative rotation and delay between camera and IMU by minimizing the rotation error between them.The convergence is verified by checking Jacobian singular value and average rotation error.The experimental results show that the proposed approach can effectively guarantee the convergence of the estimates.The initialized relative rotation accuracy is better than 2°,and the initialized relative delay accuracy is better than 5ms.(3)We derived an image point re-projection formula for rolling shutter camera.Using this formula to correct and track the feature points of the rolling shutter camera image can basically eliminate the rolling shutter effect.(4)We introduce the camera-IMU extrinsic parameters and the rolling shutter image readout time into the state space,and proposed to optimize these sensor parameters with the odometry states jointly.The system equation is dynamically adjusted according to the observability of these parameters to avoid divergence.The experimental results show that the relative rotation accuracy can reach 0.7°,the relative translation accuracy can reach 0.02 m,and the delay accuracy can reach sub-millisecond.(5)To eliminate the scale drift of low-cost IMU,five-degree-of-freedom optimization is proposed.When loop closure is detected,the scale,heading angle and three-dimensional position of each keyframe in the database are optimized.The experimental results show that the five-degree-offreedom optimization is better than the four-degree-of-freedom optimization in eliminating cumulative position errors and scale drift.2.A detailed study on models and methods for low-cost GNSS/IMU/Camera sensor fusion is performed.Owing to the inaccurate result from pseudorange positioning,the poses of keyframes are not constrained well in short term.Therefore,we proposed to constrain the system with the differential positioning result from epoch-differenced carrier phase.The experimental results show that the differential positioning accuracy can reach centimeter level.Considering the low-cost receiver has weak anti-interference ability and is easy to generate outliers,it is proposed to suppress the outliers with robust estimation and states test.The experimental results show that the proposed method can effectively reduce the influence of outlier.To fuse the information from these sensors,a pose fusion approach is proposed,where the poses from the visual-inertial system is constrained by GNSS positioning results.The experimental results show that with the assistance of GNSS,our pose fusion algorithm can further reduce the error of visual-inertial odometry/SLAM,especially the scale error.3.In order to put our approaches into practical application,we design and implement an integrated navigation platform to meet the needs for sensor data collection and processing.For time synchronization of sensor data,we use the pulse signals output by IMU and camera to determine the timestamp of IMU data and image;on the other hand,we use the UTC PPS pulses output by the GNSS receiver to align the local time with UTC.The experimental results show that the camera-IMU hardware delay fluctuation is stable within 0.15 ms,and the alignment accuracy is better than 7?s.Finally,we carried out walking and riding experiments to verify the performance of the system in the indoor and outdoor environments.We proposed to use the PPK or UWB positioning results and the pose obtained by the visual-inertial odometry/SLAM to generate the ground truth for the quantitative evaluation of the trajectory error.The experimental results show that the position error accumulation rate of the monocular visual-inertial odometry is around 0.03m/m and the heading angle error accumulation rate is around 0.008°/m under various scenarios.After the pose fusion,the position accuracy is better than 3m,and the heading angle accuracy is better than 0.12°.