Research On SLAM System Based On Stereo Vision In Indoor Dynamic Scene

In recent years,mobile service robots have gradually appeared in people’s life.Simultaneous Localization and Mapping(SLAM),one of the key technologies of robot intelligence,has become a hot topic in the field of robot research.Up to now,some scholars have proposed relatively mature Visual SLAM schemes,but the problem of poor positioning accuracy in dynamic environments has not been solved yet.In view of this,this thesis mainly studies the SLAM technology based on stereo vision in indoor dynamic scene,and proposes a SLAM system scheme suitable for dynamic environment and able to obtain semantic information of the surrounding environment.Firstly,the overall framework of this system is proposed based on ORB-SLAM2 algorithm.The process and module functions of the original ORB-SLAM2 framework are analyzed.Aiming at the three problems of poor robustness of the original framework in dynamic environment,inability to obtain semantic information of the surrounding environment and sparse point cloud map,an improved system scheme is proposed in this thesis.Aiming at the problem of large error of camera pose estimation in dynamic environment,a visual distance calculation method combining semantic information and geometric constraints to remove dynamic feature points is proposed in this thesis.To solve the problem of feature point aggregation,an adaptive threshold feature point extraction algorithm based on ORB algorithm is proposed to obtain image information comprehensively.Then,based on the semantic information of the image and L-K optical flow method,the dynamic feature points were roughly filtered out.The basic matrix was estimated by the coarse filtered image,and the polar line distance was calculated according to the matrix.Then detect the dynamic feature points again.Experimental results show that the proposed algorithm can effectively eliminate feature points on dynamic objects and improve the accuracy of pose estimation.Aiming at the problem of lack of environmental semantic information in the generated map,this thesis proposes to add an instance splitting thread for extracting semantic information based on the original ORB-SLAM2 framework.In order to ensure the real-time performance of SLAM system,YOLACT network with fast segmentation speed and high accuracy is selected to carry out semantic annotation for RGB images.The extracted results are not only used for motion consistency check,but also fused with point cloud map to generate 3D semantic map.Aiming at the problem that the original framework can only generate sparse point cloud map,this thesis proposes a 3D dense semantic map construction algorithm based on drawing key frame.The single frame point cloud is generated according to the selected drawing key frame.After the single frame point cloud is filtered,the image information and the estimated pose are used for global Mosaic to generate the 3D point cloud map.Then,the semantic information is fused with point cloud map and converted into octree map.Finally,the feasibility and effectiveness of the SLAM system algorithm designed in this thesis are verified by deploying the system and conducting experiments based on the experimental platform Ty Ran in real scenes.