Research On Semantic SLAM Method Of Mobile Robot Based On RGB-D Information

With the development of sensors and artificial intelligence technology,the tasks that mobile robots can complete are becoming more and more difficult,which requires mobile robot to be able to better perceive the surrounding environment.This promotes space perception based on RGB-D information to become a popular research direction at present.In the field of visual SLAM,RGB-D information can reduce the large amount of calculation in the localization and mapping process of mobile robot,and improve the calculation speed and accuracy of SLAM algorithm.However,the semantic information in the environment obtained through RGB-D information has not been fully utilized.In order to utilize semantic information in the environment more effectively,this paper proposes a mobile robot semantic SLAM system based on RGB-D information.While the mobile robot estimates its own motion and pose,it detects,classifies and segments objects in the environment,and then constructs a semantic map containing rich information.The main research work and conclusions of this paper are as follows:Firstly,the framework and principle of the ORB-SLAM2 algorithm with RGB-D information as input are introduced,and improved on it,through adding independent dense point cloud mapping thread,inserting dense point cloud in the tracking thread and updating the location of dense point cloud in the loopback detection thread.A visual SLAM system for dense point cloud map construction is proposed.Finally,the visual SLAM system in this paper is verified experimentally in a laboratory environment,and its performance is quantitatively and qualitatively analyzed.Secondly,we research on target detection algorithms based on deep learning,and compare different network models based on the needs of this topic in order to select YOLOv5s target detection algorithm as the method for acquiring semantic information in this paper.Then,a data set of common objects in the laboratory environment is constructed and annotated,the YOLOv5s network model is trained and tested,and its algorithm performance is analyzed.By modifying the detect.py file in the program,YOLOv5s can call the Kinect V2 camera in order to achieve the purpose of real-time detection of objects in the real environment.Thirdly,we design the 3D semantic map construction algorithm.Through Unix domain socket communication mode in this paper,the visual SLAM system with YOLOv5s fusion target detection algorithm,through the two-dimensional space to the three-dimensional space of the projection,generating a little cloud label.Through LCCP segmentation algorithm,point cloud segmentation task is completed.Then,the object model is fused and updated by means of data association,and the map information is stored in the form of octree,so as to achieve the purpose of constructing 3D semantic map.Finally,the construction method of mobile robot platform is designed from two aspects of hardware and software,and ROS system is used to realize the purpose of remote control of mobile robot walking on PC side.Then,the semantic SLAM system is deployed on mobile robot,and the semantic SLAM system is run in laboratory environment.The 3D semantic map of the surrounding environment is constructed while the pose of the mobile robot is obtained,which verifies the effectiveness of the semantic SLAM system proposed in this paper.