Research On Semantic Map Construction And Semantic Navigation Of Service Robots Based On Deep Learning

Semantic maps are the basis for the service robot to recognize the environment,which is also the premise for the robot to realize autonomous decision-making.One of the core issues in the field of robotics is how robots can use semantic information in the environment to recognize the environment like humans,build a semantic map of the environment,and use semantic information to navigate.The use of semantic information for navigation is the ultimate goal of semantic map construction,which is also an embodiment of the interaction among robots,people and the environment.This paper uses the theory of SLAM(Simultaneous Localization and Mapping)and deep learning to study the semantic map construction method and semantic navigation strategy.Firstly,this paper studies the visual SLAM algorithm which is based on the depth camera and proposes a point cloud map construction algorithm based on the improved ORB-SLAM2 algorithm.Then,we built a semantic map based on deep learning technology.Finally,robot's semantic interaction and semantic navigation are realized which is based on the semantic reasoning mechanism.Firstly,the construction method of environment map is studied from the perspective of mapping accuracy and real-time.In the research of 3D dense point cloud map construction,considering the guarantee of construction accuracy and reduction of computing resources,a depth camera is selected as the sensor for collecting environmental data in this paper.The ORB-SLAM2 algorithm is selected as the basis for the construction of 3D point cloud maps based on the comparison between theory and experiment.What's more,a method for constructing dense point clouds based on ORB-SLAM2 is proposed.In addition,considering the speed and accuracy of semantic information's insertion and updating,and the angle of reducing the memory occupied by the map,a three-dimensional occupation grid map construction method based on octree map is proposed.Secondly,a method of semantic segmentation and semantic fusion is proposed based on convolutional neural network.In the study of semantic map construction methods,this paper studies the real-time semantic segmentation technology of single frame images in the basic principles of convolutional neural networks.Aiming at the problem that the semantic and point cloud generation may not be synchronized,a method of adding temporal information to the semantic segmentation is proposed.Then,a framework of point cloud and semantic fusion is proposed based on two fusion strategies: temporal registration and highest confidence fusion.Finally,the 3D semantic map is constructed in real time at the speed of2 HZ,and the algorithm is verified using the TUM dataset.Finally,a fusion map construction method is proposed and semantic navigation is realized.In the research of fusion map construction and semantic navigation methods,considering that the most related research on semantic map construction can't realize navigation,a map representation form that fuses two-dimensional grid maps and threedimensional semantic maps is proposed.A data association method is researched to realize the synchronization and update of the data of the raster map and the semantic point cloud.Based on this,a semantic reasoning mechanism and semantic information-assisted path planning rules are proposed.It is realized that the robot can use the semantic information of the two-dimensional grid map for navigation based on natural language information such as language and text.