Research On Robot Localization And Mapping Based On Semantic Information

One of the core technologies for the intelligentization of mobile robots is the Simultaneous Localization and Mapping(SLAM)technology.Due to the rapid advancement of computer vision technology,researcher’s extensive attention has been attracted by the visual SLAM.At present,the visual SLAM system is mainly based on the research of static and simple environment,and the system performance is severely interfered in the complex environment,the positioning error of the robot will increase,and the built map will also have ghosting problems.Aiming at the positioning and mapping of robots in complex environments,this paper combines environmental semantic information to design a visual SLAM system.Aiming at the problem of semantic information extraction in indoor environment,an instance segmentation algorithm FCOS-MASK based has been designed based on target detection algorithm FCOS.In this paper,the lightweight Io T platform Mobilenet＿V2 is used in the computer algorithm to increase the speed of the feature extraction,and then a MASK branch is upgraded to carry out semantic feature extraction.Finally,the instance information in the environment is calculated by combining the results of the two branches in a linear combination.The results show that FCOS-MASK achieves the balance of speed and precision.Aiming at dynamic targets in the environment,a dynamic target detection algorithm based on sparse optical flow method and prior knowledge is proposed.The sparse optical flow method is used to predict the optical flow field..Combined with instance segmentation algorithm FCOS-MASK,instance information of dynamic objects is extracted to determine the dynamic objects as prior knowledge.In this paper,on the basis of ORB-SLAM2 was improved,the accuracy of positioning in the dynamic environment has been improved through the elimination of dynamic feature points,and the problem of ghosting in dynamic environment is solved.In this paper,establish an indoor semantic map is investigated systematically.Segmentation algorithm based on instance FCOS-MASK network for instance segmentation,and connecting with the depth of information filtering,point cloud segmentation complete two-dimensional semantic point cloud to semantic point cloud three-dimensional map of transformation,and to establish semantic information database,the semantic information of indoor environment has been perceived.the semantic information of indoor environment.On the Turtlebot2 mobile wheeled robot equipped with Kinect V2,positioning accuracy and mapping quality comparison experiments were carried out with ORBSLAM2 system in real environment and dataset environment.The results showed that the positioning accuracy of the SLAM system designed was greatly improved compared with ORB-SLAM2,and the system includes the ability to build three-dimensional semantic graph.