Research On Visual SLAM Of Mobile Robots In Dynamic Scenes

Simultaneous Localization And Mapping(SLAM)is an important means for mobile robots to realize autonomous motion.Therefore,the research on SLAM is extremely important.At present,most existing SLAM systems assume that the environment is static.However,there are abundant dynamic scenes in real life,as a result,the existing SLAM system has low accuracy in positioning and mapping.In response to the above problems,the thesis carried out the research work of the visual SLAM system of mobile robots in dynamic scenes.The main research contents of the thesis are as follows:Firstly,according to the applicable scenarios of various cameras,the camera type used in SLAM in this article is selected.Then,the mathematical model and distortion model of camera imaging were constructed.Finally,the characteristics of various calibr ation algorithms are analyzed,and one of the calibration algorithms is selected to calib rate the selected camera.The ORB-SLAM2 algorithm framework is analyzed,and its s hortcomings in positioning and mapping in dynamic scenarios are summarized,and an improved system framework is proposed.Secondly,in order to solve the problem that the point feature extraction algorithm is prone to local aggregation,the quadtree algorithm is used to make the feature distribution uniform.Aiming at the problem that after removing the features of dynamic targets in dynamic scenes,the number of SLAM extracted point features will decrease and the point features of weak texture regions will be missing,a point-line fusion scheme is proposed,which improves the accuracy and robustness of pose estimation in dynamic scenes.In the visual SLAM front-end,the YOLO v4 network is used to detect dynamic targets in the image,and then the dynamic features in the scene are filtered out,and with the help of the deduced reprojection error model based on point and line features,eliminates the pose estimation error of the SLAM system in dynamic scenarios.Then,aiming at the local optimization algorithm can not well eliminate the problem of the accumulation of front-end pose estimation errors,researched the loop detection optimization algorithm based on point and line features,and with the help of the pictures in the public data set and the dictionary based on point and line features obtained by offline training,the effectiveness of the algorithm is verified.Aiming at the ghosting problem of mapping in dynamic scenes,a key frame strategy for mapping is proposed,and mapping experiments are carried out based on public data sets.The experimental results show that using the key frames proposed in this article for mapping can construct a point cloud map and an octree map of the scene without ghosting.Finally,the SLAM system in this thesis is compared with the ORB-SLAM2 system in positioning experiments,the experimental results show that the SLAM system in this article has higher positioning accuracy.Design experiments to compare the real-time performance of this article's SLAM system and Dyna SLAM system,the experimental results show that the SLAM system developed in this article has better real-time performance.Based on the robot platform,the positioning and mapping experiments in the real environment are carried out,the experimental results show that the algorithm in this thesis has higher robustness and better mapping effect.