Research On Semantic SLAM Based On Uncertainty Model And Relocalization

The localization and mapping of robots is crucial for indoor service robots and is a prerequisite for follow-up robot navigation and environmental awareness.Accurate positioning ability can be used to complete the navigation requirements more accurately and accurately.It can be used for efficient real-time construction of maps to better help robots extract the semantic information of 3D scenes and complete the recognition of the environment.This topic focuses on mobile robots.There are three sub-topics of feature point uncertainty model,relocalization algorithm,and semantic construction in positioning and mapping.The purpose is to obtain a highly accurate and robust SLAM system by improving the feature uncertainty model and the relocalization algorithm,and the improved SLAM system can obtain high-precision three-dimensional semantic scenes.This subject first analyzes the error measurement model of RGBD camera.From a large number of experiments,the relationship between the number of layers in the scale space and the uncertainty model is summarized.By establishing new uncertainties and improving the accuracy of the system operation,the scale-based uncertainty model is associated with the scale and depth parallax of feature points,and the effect of camera motion patterns on the uncertainty model is analyzed through experiments.Then this topic proposes a global hierarchical registration localization algorithm based on point cloud aiming at some problems in the actual situation,which solves the problems of robot booting,restarting auto-bit,relocalization path dependence and so on.relocalization,as an important part of SLAM technology,is of great help to improve the robustness of the robot.In this paper,it is possible to solve large-scale point cloud registration real-time solving pose through the hierarchical processing of large point clouds and the dimensionality reduction of poses.The proposed pose consistency detection technology ensures the accuracy of relocalization.And using the neural network to complete the fusion of multi-dimensional pose information,to ensure the accuracy of the pose.The relocalization algorithm can greatly improve the recall rate of the algorithm under the premise of guaranteeing accuracy,and effectively realizes the multiplexing of the map by the robot.Finally,in order to realize the perception of the environment of indoor service robots,this topic uses the current best case segmentation network model Mask R-CNN for semantic sensing,and semantically perceives the key frames generated by the front end of the SLAM system.Interested key objects are semantically segmented and rendered.Finally,the three-dimensional semantic scene is established by the improved camera coordinates provided by the SLAM system and the proposed semantic voxel filter.In a word,on the basis of the traditional SLAM architecture,this paper aims at the existing problems of the existing algorithms and improves the two sub-links in the system: uncertainty model and relocalization.The improved SLAM system has both accuracy and robustness.With certain improvements,the instance-division network was finally embedded in the improved SLAM system to complete the establishment of a real-time semantic SLAM system.