Researh On 3D Semantic Map Construction Based On Edge Detection

Visual SLAM(Simultaneous Localization And Mapping)is a key technology for intelligent mobile robots to realize precise Localization by sensing unknown surroundings.Based on the existing research,how to construct semantic map using semantic information in the environment and use it for navigation has become a hot issue.In the mainstream semantic SLAM system,the convolution neural network is used for object detection or semantic segmentation in 2D image to obtain the semantic label of the object,and then it is converted into the corresponding 3D space by means of semantic mapping,to obtain the 3D point cloud with semantic information.In this way,the segmentation object is not clear,the edge of the object is fuzzy,the semantic information is easy to be lost and the constructed map is incomplete.Therefore,in order to improve the overall accuracy and localization robustness of mobile robots in semantic map construction,a 3d semantic map construction method integrating edge detection is studied.The constructed map can make the intelligent robot better perceive the objects in the indoor scene and the overall environment,and complete its own positioning and navigation requirements,which has important application value and significance.The main work and achievements of this paper are as follows:(1)Firstly,the visual SLAM system is studied to meet the needs of intelligent robot to achieve accurate positioning and build a map suitable for indoor environment.The framework of visual SLAM algorithm and the principle of each module are briefly described.As for the selection of map construction algorithms,the current mainstream algorithms are compared and classified,and the ORB-SLAM2 algorithm is selected as the basic framework of semantic map construction by analyzing the basic theories and advantages and disadvantages of the algorithms.(2)Secondly,aiming at the limitations of semantic SLAM in semantic recognition of objects in 3d space,a point cloud semantic segmentation algorithm integrating edge detection is proposed by convolution of 3D point cloud directly.Based on the KPConv deep learning network of point cloud,the semantic edge detection network is integrated,and the attention mechanism is introduced in the semantic edge detection network,which can effectively extract the edge information of point cloud and generate precise semantic edge.A fusion module is designed to effectively fuse edge feature results from semantic edge detection network and region feature results from semantic segmentation network,and further refine the fusion results.A double semantic loss function is used to obtain better semantic boundary results.After fusion,the network improves the disadvantages of unclear edge contour and unclear segmentation,and has better segmentation performance for small objects.The semantic information obtained is used in the point cloud map constructed by the visual SLAM system to realize the updating and fusion of semantic map.(3)Finally,the paper studies the construction of dense point cloud map and semantic point cloud map,including the construction process and overall system design of point cloud semantic map,the construction method of dense point cloud map,and the construction of point cloud semantic map through semantic association and update.The validity of the system is verified by using open data sets,and the accuracy of the semantic segmentation algorithm based on fusion edge detection for 3D point cloud is verified,and the dense point cloud semantic map has better readability compared with the map constructed by traditional visual SLAM system.(4)In addition,in order to further utilize semantic map and reduce storage space,semantic octree map is studied,and a map suitable for indoor scene positioning and navigation and semantic perception of service robots is constructed.