Research On Real-Time Positioning Method Based On Information Fusion Of Vision And Inertial Navigation

Autonomous unmanned equipment such as smart mobile robots and drones are improving people’s production and lifestyles to a large extent.A reliable positioning system is required in these devices.The GPS(Global Positioning System)method can be used for positioning in an outdoor environment,but this method is powerless in a scene where the GPS signal is weak,such as indoors.At the same time,the Simultaneous Localization and Mapping(SLAM)method only needs to process the data collected by the self-carrying sensor to estimate its own position and reconstruct the surrounding scene in an unknown environment.The traditional SLAM method is mostly based on a single vision sensor,so it is not robust in scenes such as fast motion and low texture,and there is also the problem of uncertain scale.Therefore,it is necessary to study a real-time positioning method based on the fusion of different sensor information.The main work of the thesis includes:(1)Aiming at the disadvantages of single vision sensor in terms of stability and robustness,IMU(Inertial Measurement Unit)sensor is introduced to obtain VI-SLAM(Visual Inertial SLAM)system of visual and inertial information fusion.Based on the maximum a posteriori estimation method,the unified optimization objective function form of the two sensors in the VI-SLAM system is deduced,so that the two are tightly coupled together by a nonlinear optimization method.In addition,a parameterization strategy for adaptive selection of landmark points is added to enhance the adaptability of the algorithm in large-scene environments.(2)In view of the current low efficiency of IMU initialization,the initialization method of the VI-SLAM system is improved,and a two-stage initialization method of "loose-tight"coupling is designed.The loose coupling stage uses the visual estimation results to initialize the IMU parameters,and then the tight coupling stage performs visual-inertial joint optimization.The experimental results on the accuracy and efficiency of the initialization method on the public data set show that the average initialization scale error is about 10%,and the additional time related to IMU initialization is less than 200ms,which proves the practicability of the system initialization method.(3)The proposed algorithm was tested on the public EuRoC data set and compared with the other two excellent algorithms.The experimental results show that the positioning accuracy of the algorithm proposed in this thesis is within 2cm in the simple sequence and within 8cm in the difficult sequence,which is better than the other two algorithms in terms of accuracy.In terms of efficiency,it can process about 23 frames per second,meeting real-time requirements.In addition,in order to verify its practicability and stability,the algorithm was tested in actual scenes.The test scenes included a stairway environment with large jitter and an outdoor campus environment.In the process of going up and down the stairs,the algorithm can run stably without tracking failure.The estimated height difference of the left and right stairs is within 0.5m.In the outdoor environment experiment with a track length of about 580m,the algorithm also runs stably.The positioning error in the closed path from the start point to the end point is about 2.5m,and the relative error is within 2%,which verifies the effectiveness of the improved algorithm for large-scene environments.