| In recent years,simultaneous localization and mapping(Simultaneous Localization and Mapping,SLAM)technology has gradually become a hot topic of research.The traditional SLAM system based on visual point features will cause tracking failure or inability to track due to insufficient feature point extraction in weak texture and structured scenes.In addition,when the robot moves too fast indoors,tracking failures can also occur.In order to solve the above problems,this thesis introduces higher-dimensional line features as a supplement on the basis of point features,and combines inertial measurement units(IMU)to propose an SLAM system based on the fusion of visual point-line features and IMU,which can be applied to navigation and positioning in indoor structured scenes or weak texture scenes.The main research contents and achievements of this thesis are as follows:(1)In the front-end visual odometry,aiming at the problems of ORB feature point extraction algorithm overemphasizing the prominent texture,resulting in increasing the difficulty and time-consuming matching,a quadtree based on adaptive depth division is proposed to realize the homogeneous extraction of feature points,and has been experimentally verified.Based on the point feature,the line feature is introduced,and for the splitting phenomenon of the line segment extracted by the LSD(Line Segment Detector)algorithm,this thesis proposes a line segment detection algorithm based on multi-scale and multi-resolution,and verifies through experiments,the improved line feature extraction algorithm effectively overcomes the shortcomings of the original LSD,taking into account the details of the line segment in the environment and the length of the line segment.For the matching of point and line features,the BRIEF(Binary Robust Independent Elementary Features)descriptor and LBD(Line Band Descriptor)descriptors are used to describe and match the point features and line features,and after the matching is completed,the RANSAC(RANdom SAmple Consensus)method is used to mismatch and eliminate the point features.(2)In the back-end optimization part,the visual and IMU information are fused by the tight coupling scheme,and the point feature reprojection error,line feature reprojection error,IMU pre-integration error,a priori constraint caused by marginalization and constraint caused by loopback detection constitute the overall objective function,and the optimization is completed through the sliding window strategy,in which for each error,the Soft One robust kernel function is introduced to ensure the optimization accuracy while improving the running speed of the system.Since the back-end optimization and loopback detection are processed for keyframes,the selection of keyframes is crucial,and a keyframe filtering strategy is proposed based on this.Experimental verification of the above improvements shows that the operating speed of the system is increased by an average of 684 ms.(3)The SLAM system based on the fusion of visual point and line features and IMU is designed and realized,and the experimental comparison with two advanced algorithms on the public dataset shows that the proposed system has higher positioning accuracy and good robustness.The system is transplanted to the built mobile robot embedded platform,and experimental verification is carried out for structured and weak texture scenarios,and the experimental results show that the algorithm studied in this thesis can run on the embedded platform in real time and has good positioning accuracy. |
