Construction Of Indoor Semantic Map Based On Visual SLAM

With the development of science and technology and the improvement of living standards,robots have gradually entered people's daily lives.However,current home service robot usually build sparse road sign maps containing geometric information based on Simultaneous Localization and Mapping(SLAM)technology,which can only be used to perform navigation and localization tasks.In this case,the robot cannot understand the high-level semantic information of the item in the environment,and it is difficult to execute the command containing the high-level semantic information issued by the human.Under this requirement,combining SLAM and object detection and recognition technology to build a semantic map containing object semantic information has become an effective solution.The construction of indoor semantic maps begins with the construction of object-oriented se-mantic maps.Image segmentation was performed based on the YOLACT(You Only Look At Coef-ficienTs)network,and an adaptive region growth algorithm combined with depth information was proposed to optimize the segmentation of the instance,resulting in more accurate three-dimensional semantic segmentation and the semantic labeling of a single frame of image.For the object in-stances of the map,the semantic features extracted by YOLACT and the color features of the LAB space were used to characterize the appearance of the object instances.By combining appearance consistency and spatial consistency,instance matching was performed to implement the object con-struction and object update to build an object database.A semantic map containing object instance information is constructed based on the octree data structure.This thesis also studies the construction of semantic maps for dynamic environment.Aiming at the dynamic targets in the environment,this thesis proposes a dynamic target detection algorithm based on optical flow and prior knowledge.The FlowNet 2.0 network is used to predict the optical flow field,and the result of YOLACT semantic recognition is used to fuse the prior knowledge of the dynamic properties of the object to determine whether it is a dynamic object.Based on ORB-SLAM2,this thesis improves the localization accuracy in dynamic environment by eliminating dynamic feature points,thus improving the construction performance of semantic map in dynamic environment.Based on the Robot Operating System(ROS)platform,this thesis conducts experiments on both object-oriented semantic map construction and semantic map construction for dynamic en-vironment.In this thesis,both object-oriented semantic map construction and semantic map con-struction for dynamic environment are carried out in public data sets and in real-world scenarios of simulated indoor home environment established in the laboratory.Experiments show that the object-oriented semantic map construction algorithm proposed in this thesis can identify and match object instances in the environment to maintain object instance information in the semantic map.And the semantic map for the dynamic environment proposed in this thesis can realize the accurate localization of the robot in the dynamic environment,and improve the semantic map construction performance in the dynamic environment,which verify the feasibility and robustness of the algo-rithm in this thesis.