Research On IMU-assisted Stereo Vision SLAM Method

Simultaneous localization and mapping(SLAM)based on multi-sensor fusion has the advantages of high precision,strong robustness and good consistency.It is a research hotspot in the field of SLAM.Inertial measurement unit(IMU)has high positioning accuracy in a short time,but there is a drift problem for a long time.The information acquired by the vision sensor is rich.There is no accumulative error,but the positioning cannot be achieved in the case of occlusion,blur or fast motion.Fusing IMU with vision can overcome the shortcomings and achieve high-precision SLAM.Focus on the method,in order to further reduce accumulative error in the process of data association,the visual loop closure detection method based on DBoW3 is introduced into the fused SLAM.Moreover,the marginalization processing is improved.Due to the introduction of loop closure detection and the optimization of data association,the precision of fused SLAM is greatly improved.The work of the thesis mainly includes:(1)Based on the pinhole camera and IMU error model,the positioning principle of the fusion of IMU and stereo vision is summarized.Based on the principle,an improved tight coupling inertial vision fusion method is proposed.The method uses the nonlinear optimization method to optimize the IMU error term and the visual re-projection error term,and adopts an improved marginalization strategy in local optimization.(2)In order to further improve the consistency and accuracy of SLAM,DBoW3-based visual loop closure detection is studied and introduced into IMU and stereo vision fused SLAM.In the DBoW3-based visual loop closure detection method,a visual dictionary tree is established by extracting oriented fast and rotated brief(ORB)features from keyframes,and a suitable loop closure detection threshold is determined through a large number of experiments.On the basis of the loop closure detection results,the global optimization method based on Ceres is adopted to re-linearize past states,which reduces the accumulative error in the process of data association.(3)For the improved fused SLAM method proposed in the thesis,a lot of experiments have been carried out on the Euroc standard datasets and datasets collected by experimental team.The datasets include:Euroc datasets,campus datasets and indoor datasets.There are two kinds of experiments in this thesis.One is the comparison of the accuraey of the improved fusion method with the pure visual method,and the other is the accuracy comparison between the method of adding loop closure detection and the original method.The comparison of precision results utilizing APE indicator shows that the improved method is more accurate.This thesis contains 43 figures,20 tables and 64 references.