Research On SLAM System Of Point And Line Features Fusion Based On Fisheye Camera

SLAM(Simultaneous Localization and Mapping)is one of the key technologies in the field of unmanned driving.Its purpose is to obtain the vehicle location information according to the surrounding environment information.Vision and inertial information fusion is one of the main research directions in SLAM.It combines the advantages of visual sensor and inertial sensor,and fuses the information of the two sensors.The scale factor of monocular vision is determined by the motion information of inertial sensor IMU(Inertial Measurement Unit),and the drift of IMU is constrained by the environmental information obtained by camera.The current VI-SLAM(Visual-Inertial Simultaneous Localization and Mapping)system has the following problems: in the case of sparse texture,fast-moving,sudden light changes,the lack of feature points will lead to reduced system positioning accuracy and even system failure.In order to improve the robustness of the system in the above cases,this paper will use the fisheye camera with large FOV(Field of View)to improve the existing VI-SLAM system with point and line features fusion.In this paper,the SLAM system based on vision inertial fusion is taken as the research object,and the effectiveness of point and line features fusion in VI-SLAM system is studied.This paper mainly discusses the sensor models and their calibration,the line feature correlation algorithm in VI-SLAM system,and the VI-SLAM system based on point and line features fusion.For the part of sensor model,the projection and distortion models of fisheye camera are analyzed,and the working principle and error model of inertial measurement unit are described.The fisheye camera and IMU’s reliable internal and external parameters are obtained by joint calibration.The front end of the traditional VI-SLAM system uses feature points or optical flow to solve the pose.On this basis,this paper introduces the line feature in space,and interprets the line feature extraction algorithm and descriptor in detail.In order to improve the computational efficiency of line features in space in the system and solve the problem of over-degree of freedom in back-end optimization,the Plucker Coordinates and orthogonal representation are introduced.The comparison experiment of line feature extraction by collecting images proves the superiority of the line feature extraction algorithm used in this paper.Then,the whole VI-SLAM system framework is introduced.Firstly,the visual initialization is completed,and the data of visual observation and IMU observation are aligned to restore the real scale.Then,IMU pre-integration model and point and line feature observation model are derived,and their Jacobian matrices are calculated respectively.Back-end optimization uses Bundle Adjustment based on sliding window to optimize the visual reprojection error,IMU integral error and prior error.Finally,in order to verify the positioning accuracy and computational efficiency of the program designed in this paper,the accuracy comparison experiment and frame rate comparison experiment are carried out on the public data sets.The experimental results show that the root mean square value of absolute pose error of the program designed in this paper is 11.5% lower than that of VINS-Mono,and the relative pose error is 66.8%lower than that of VINS-Mono.Compared with PL-VIO,which also uses point and line features,the program designed in this paper can estimate the pose in real time.The frame rate can reach 10 Hz.By setting up a mobile experiment platform,the positioning experiment has been carried out in complex indoor scenes.The system can still estimate the pose accurately when working in strong exposure,weak illumination and sparse texture experience.The experimental results showed that program designed in this paper can estimate the pose accurately and robustly.