Research On Dynamic SLAM Based On Vision And Inertial Information

In recent years,China's independent communication technology has been breaking through industry barriers,especially the popularity of 5G communication technology,which has further promoted the development of autonomous motion robots.The outbreak of COVID-19 requires a reduction of close human contact,and a large number of service robots have been put into the front line of the epidemic,playing an important role in supplies delivery,area disinfection and other aspects.Simultaneous Localization and Mapping(SLAM)technology is the main method to realize autonomous motion of mobile robots in unknown environments.However,in the actual use process,these service robots generally have problems such as low working efficiency and poor positioning accuracy in complex scenes.This is because traditional SLAM algorithms are mostly designed for static environments,leading to poor positioning performance of mobile robots in complex dynamic scenes,which largely limits the application of SLAM technology in real scenes.This dissertation mainly studies how to improve the robot's perception ability of dynamic information in the real scene,so as to realize the precise positioning and smooth operation of the robot in the dynamic scene.The main work of this dissertation is as follows:Firstly,the basic model of visual inertia SLAM is studied in this dissertation,and the monocular camera model,distortion model and visual reprojection error model are discussed in detail.The motion model of the inertial sensor is studied in depth,and the formula of IMU pre-integral is deduced in detail.Secondly,this dissertation studies the initialization and feature extraction methods of visual inertial system.Aiming at the feature points extracted by traditional ORB feature extraction methods under dynamic environment contain a large number of dynamic features,which leads to the problem of feature mismatching,a dynamic feature point detection algorithm based on polar geometry is proposed.In order to solve the time-consuming problem of ORB feature extraction,sparse optical flow method is used to track features in non-key frames.Aiming at the problem of easy tracking loss in dynamic environment,this dissertation carries out relocation according to the local map database and IMU data maintained.The localization accuracy and running speed of the improved SLAM system are significantly improved in the dynamic environment.Then,in this dissertation,the dynamic SLAM the back-end optimization part of the system are introduced in detail,within the sliding window will be in the form of tightly coupled visual and inertial information fusion,according to dynamic environment disturbance to the visual SLAM problem,put forward an adaptive adjustment strategy,this strategy can be adjusted according to the environmental situation of back-end optimization in the process of inertia weight of the information and visual information,improved SLAM robustness of the system in a dynamic environment.Finally,the accuracy and robustness of the proposed algorithm in dynamic environment are verified by experiments.In this dissertation,the performance of TUM-VI visual inertia dataset is tested on static and dynamic sequences respectively,and compared with WINS-MONO and ORB-SLAM2 algorithms under the same experimental conditions.The final experimental results show that the improved SLAM system in this dissertation can run smoothly in the dynamic environment.Compared with the ORB-SLAM2 algorithm,the processing time of each frame of the front-end oemometer in this dissertation is reduced by72.82% in the static scene,and by 76.04% in the dynamic environment.Compared with the VINS-Mono algorithm,the positioning accuracy of the proposed algorithm in dynamic environment is improved by about 47%.