Research On Filter Based Robust SLAM Algorithm

Simultaneous Localization and Mapping(SLAM)is widely used in field of robotics,including mobile robot,automatic vehicle,etc.In recent years,multi-sensor fusion,with the advantages of high accuracy and good robustness,has received extensive attention among academic and industrial field,and is applied to SLAM(e.g.Visual-Inertial Odometry,VIO).However,the majority of SLAM-related algorithms is based on the assumption that the observation noise of the system is Gaussian distribution.Actually,in the real world,sensors might be affected by special environment and sometimes even sensors themselves might get into problems.In these cases,observation data might have many unrealiable outliers,and for this reason the observation noise is non-Gaussian distribution.In this thesis,a variational Baysian adaptive SLAM algorithm is proposed to increase the robustness and location accuracy of the system.In this algorithm,the covariance of observation noise is modeled as inverse-Wishart distribution,so that the observation noise is non-Gaussian model.Besides,cubature Kalman filter is applied to the system to handle the problem related to the nonlinear system.According to extensive simulation experiments and real world dataset experiments,we validate the effectiveness of our proposed 2-D SLAM algorithm.Based on the above study,the proposed robust SLAM algorithm is adapted to a visual-inertial odometry system to obtain a filter-based tightly-coupled fusion algorithm.Moreover,unscented transformation is introduced to the system to compute Jacobian matrices.The effectiveness of our proposed 3-D VIO algorithm is validated based on opensource Eu Roc dataset.Lastly,an unmanned aerial vehicle platform is built to record visual and inertial dataset,and effectiveness of proposed VIO algorithm is tested on the dataset.