Narrow Space Navigation Method Combining Vision And Inertial Measurement Unit

Narrow spaces such as mine roadways,tunnels and subway passages are prone to accumulate toxic and harmful gases,stagnant water and falling rocks,etc.,which endangers the safety of personnel and normal production and life,and requires daily inspections.Among them,manual inspection has low efficiency and strong subjectivity.The judgment of risk factors depends on experience,lack of scientific,objective and accurate judgment basis and other limitations.It can no longer meet the production and development needs of digital society.It is necessary to develop independent robot platforms to replace humans.Related work,which is of great significance for safe production and personnel safety,as well as intelligent production and digital society.This thesis uses Simultaneous Localization and Mapping(SLAM)algorithm and visual navigation algorithm,based on the ORB-SLAM2 framework,fusion of visual and Inertial Measurement Unit(IMU)sensors,completed the following work:(1)Compare and determine the feature detection algorithm suitable for narrow space environmentFeature detection algorithms are mostly used in feature-rich indoor and common open outdoor scenes.In actual narrow spaces,there is uncertainty in the performance of feature detection algorithms.For specific environments in narrow spaces,research includes SIFT,BRIEF,Shi-Tomasi and other feature detection and description extraction algorithms Combined with experiments to compare and analyze the feature detection quantity,time consumption,algorithm robustness and other aspects to determine the most suitable feature detection algorithm for the narrow space environment.(2)Improved feature mismatch elimination algorithmCorrect feature matching is the key to estimating camera pose in visual SLAM.Aiming at the shortcomings of Random Sampling Consensus(RANSAC)in the existing feature mismatching elimination algorithm in terms of efficiency,accuracy,and existence of degradation,it is researched and improved to Progressive Sample Consensus(PROSAC).Experiments based on simulation data and real data respectively verify the feasibility and effectiveness of the improved PROSAC algorithm.(3)Realization of SLAM system integrating vision and inertial navigationIn view of the lack of scale information in the pure visual ORB-SLAM2 system and the inability to cope with the rapid movement of the camera,the theories of IMU pre-integration,multi-sensor calibration and fusion are studied.Based on the ORB-SLAM2 framework Through the loosely coupled framework,fusion of visual and inertial unit data in the initialization stage achieves calibration coefficients such as scale coefficients and IMU offsets.And the tightly coupled framework fuses visual and IMU data to estimate the state of the system and build a hardware system that combines visual and inertial navigation.Experiments were conducted on public datasets and narrow spatial datasets,respectively,to verify the accuracy and robustness of the SLAM system combining vision and IMU.(4)Research on visual navigation methods in narrow spacesCombined with the specific narrow space environment,study the image segmentation algorithm and its characteristics,extract the structured information in the narrow space environment image,propose a visual navigation method,solve the yaw angle of the drone relative to the confined space to provide attitude feedback for the flight controller.Combined with the experiment,the feasibility of the visual navigation method is verified.The method in this thesis can realize positioning and map construction in a narrow space environment,recover scale information on motion trajectories and maps,and provide attitude feedback for UAV platforms.Through the fusion of vision and IMU,the positioning accuracy and robustness are improved.The thesis has 79 graphs,9 tables,and 92 references.