Research On Indoor Semantic SLAM Algorithm Based On Fusion Of Visual And Inertial Information

In recent years,with the development of Simultaneous Localization and Mapping(SLAM)technology,it is widely used in various life and business fields.However,SLAM technology faces various challenges in the process of its popularity and development.For example,there are often dynamic environments in indoor scenes,and dynamic targets can make it difficult for localization algorithms to achieve correct feature matching.In addition,pure visual odometry cannot perform accurate pose estimation when facing problems such as insufficient light and fast motion.To this end,this project will investigate robust and accurate localization methods for mobile robots in indoor environments,introduce semantic information using semantic segmentation methods,and reject dynamic targets by semantic information combined with multi-view geometry to improve the response capability of the localization system when facing dynamic environments.At the same time,a multi-sensor fusion approach is further investigated to improve the localization accuracy and robustness of mobile robots by using inertial sensors to compensate for the shortage of visual sensors.Firstly,in this thesis,semantic information in image frames is obtained through semantic segmentation network and used for dynamic target capture.The effectiveness of the semantic segmentation network is verified by conducting experiments on two publicly available datasets as well as a set of realistic scene data,which provides reliable support for the subsequent research.Secondly,the semantic information combined with multi-view geometry is used to achieve dynamic target rejection and avoid feature points on dynamic objects interfering with feature matching and positional estimation.After optimizing the feature extraction method,a matching algorithm based on grid motion statistics is used in this thesis to reduce the mis-matching of ORB feature points,and the algorithm is validated on a publicly available dataset.By comparing and analyzing the posture estimation effect before and after the algorithm improvement,it is verified that the improved method in this thesis is effective in improving the localization accuracy of the mobile robot in the face of dynamic environment.Then,the pre-integration of inertial sensors is investigated in this thesis,and the preintegration is used to provide more reliable initial poses for the localization algorithm.In order to further involve the inertial sensor in the back-end optimization to realize the tight coupling with the vision sensor,this thesis establishes the least squares problem using the residuals and derives the information matrix and Jacobi matrix about the inertial parameters to finally realize the tight coupling between the inertial sensor and the vision sensor.Finally,a hardware verification platform based on a quadruped mobile robot is built,and comparison experiments are designed on public data sets and realistic scenarios to verify the effectiveness of the inertial localization algorithm combined with semantic information in the face of highly dynamic environments.After three sets of comparison experiments,it can be seen that the absolute trajectory error of the method studied in this thesis decreases by about 0.5m compared to the localization algorithm using only visual point feature information.It is verified that the semantic information and inertial sensorbased localization method studied in this thesis can improve the localization accuracy and system robustness in indoor dynamic environments.