Localization And Mapping Of Indoor Dynamic Scene Based On Instance Segmentation

In recent years,with the rapid improvement of my country’s scientific and technological innovation strength,intelligent robots have been integrated into all aspects of people’s production and life.Simultaneous Localization And Mapping(SLAM)technology has also been achieved as a core technology in the field of intelligent service robots.With unprecedented breakthroughs and developments,it has gradually attracted the attention of researchers from all walks of life.Therefore,the research work of visual SLAM technology has very important practical significance and is closely related to people’s daily production and life.This paper chose the visual positioning and mapping of mobile robots in indoor dynamic scenes as the research background.Aiming at the problem that dynamic objects in the environment affect the accuracy of camera trajectory estimation,the SLAM system in this paper first extracted ORB features from the image sequence obtained by the RGB-D camera Point,used the Mask R-CNN instance segmentation network to extract the dynamic object information in the indoor scene,generated the corresponding dynamic object mask,and then eliminated the unstable ORB feature points distributed in the prior dynamic object.Aimed at the problem that there may be many mismatches in the ORB feature matching process,in this paper,the SLAM system used the Grid-based Motion Statistics(GMS)algorithm to eliminate mismatches and improve the matching quality of ORB feature points.Finally,iterative closest point(Iterative The Closest Point(ICP)algorithm calculated the camera pose and obtains a more accurate estimated trajectory of the camera,effectively avoiding the influence of dynamic objects on the camera pose calculation.The experimental results tested on the public data set show that the SLAM system in this paper has different degrees in the absolute trajectory error and relative pose error of the camera pose estimation compared with the ORB-SLAM2 system,ORB-SLAM2+YOLOv4 system and Dyna SLAM system.The decline of the experiment results expected in this article.In terms of map construction,the SLAM system in this paper first used background repair technology to fill the area occluded by dynamic objects,and then according to the RGB-D camera parameters and the estimated camera pose,the two-dimensional image processed by the Mask R-CNN instance segmentation network was processed.The image information was projected into the three-dimensional space used point cloud stitching and filtering technology to obtain an instance-level static dense point cloud map without dynamic object interference,which improved the robot’s ability to understand and perceive the surrounding environment.Finally,a semantic octree map is established on the basis of the instance-level static dense point cloud map.The experimental test results on the public data set showed that the semantic octree map could save a lot of memory space compared to the instance-level static dense point cloud map,and the semantic octree map could be applied to the navigation of mobile robots to prepare for subsequent robots to perform advanced tasks.