Research On Vision-based Indoor ORB-ISM Simultaneous Localization And Mapping

Simultaneous Localization and Mapping(SLAM),an important branch of intelligent mobile robots,plays a decisive role whether robots will complete autonomy in the exploration of unknown environments.In particular,Simultaneous positioning and mapping based on vision has always been a research hot spot in the field.Based on shortcomings existing in mobile robot vision SLAM system,including:(1)The existing online construction method based on visual maps is mostly sparse point cloud map,which cannot be used for robot navigation and obstacle avoidance;(2)In the conventional Monte Carlo positioning method,the sample degradation may occur in the sequential importance sampling;(3)Initialization problems in visual inertial fusion technology.Specifically,research from the following aspects:1.An online construction algorithm of grid map based on ORB-ISM is proposed.Aiming at problem(1),an online construction algorithm of grid map basedon ORB-SLAM2 is proposed.First,an Inverse Sensor Model(ISM)for visual SLAM is established.Secondly,for the ISM,the construction mechanism of the grid map algorithm is rearranged and detailed derivation is carried out.Finally,the specific implementation of ORB-SLAM2 grid map construction algorithm is designed,and the ISM model and the grid map model are analyzed to find the best model parameters.2.Adaptive Gradient Refined Stratified Sampling(AGRSS)is proposed.For problem(2),an adaptive gradient refined stratified sampling method is proposed.In this part,Bayes filtering algorithm and Monte Carlo algorithm are expounded firstly.Then the defects of sequential importance sampling and uniform layered sampling are analyzed.Finally,through argumentation,AGRSS can perform the refinement of sequential space,achieving a near optimal stratified sample design,the feasibility of AGRSS sampling method is established through analysis.3.Gyro Differential Calibration(GDC)was proposed.For the problem(3),GDC is proposed to calibrate the gyro bias-the gyro bias is calibrated by the visual and inertial rotation variable difference method.Simultaneously,in order to overcome the problem of single application scenario,the method of automatically selecting the Fundamental matrix or the Homography matrix is used to estimate the camera rotation variable,and the translation vector is obtained by the minimum error method.Experiments demonstrates that the monocular camera and the RGB-Ddepth camera are used to construct the grid map online.Both of them can achieve the online construction of the grid map,indicating that the ORB-ISM grid map online construction algorithm supports both monocular and RGB-D modes,and have a strong applicability.Maps constructed with RGB-D cameras in both indoor and corridor environments can clearly display the position of obstacles in the environment,validating the effectiveness of the ORB-ISM grid map construction algorithm.It is proved by experiments that AGRSS can perform the refinement of sequential space,achieving a near optimal stratified sample design,and overcoming the limitations of possible sample degradation in sequential importance sampling.The average RMSE of improved Monte Carlo localization algorithm is reduced by 5.1545;Experiments of visual inertia initialization algorithm show that the GDC method improves the calibration accuracy of the gyroscope bias,overcome the arbitrariness of monocular visual scale,and the average RMSE of the estimated trajectory of this paper is reduced by 0.0891 m,and it performs well in different scenarios.