Research On Visual Positioning And Semantic Mapping Technology In Dynamic Environments

Vision-based simultaneous localization and mapping technology is the key technology for robot localization and three-dimensional map construction.This theis focuses on the study of visual location based on static region extraction and the construction of 3D semantic map.The main research contents are as follows:(1)A visual SLAM algorithm based on the geometric consistency of the environment is proposed.The dynamic objects are segmented and removed from the static background based on the geometric consistency of the static pixels observed from different perspectives,and the feature points in the static background are extracted during the pose calculation process to estimate the camera trajectory.Combined with the improved loop detection scheme,the system accuracy is improved by reducing the weight of dynamic objects in the scene and removing specific dynamic objects.Experiments show that the proposed algorithm can effectively improve the system accuracy under dynamic scenes and provide more stable and accurate visual positioning performance.(2)A scene segmentation algorithm based on plane structure information and local convexity is proposed.First of all,the point cloud is converted into patch adjacency graph by supervoxel clustering.Secondly,the reference plane set is extracted considering the geometric topological characteristics and plane fitting degree.Then the patch set is classified initially based on the reference plane attribution of the point cloud.Finally,the local convexity feature is employed to segment the remaining patches in the non-planar area.Experiments show that the algorithm has strong anti-noise ability and effectively realizes the point cloud segmentation of 3D scenes.(3)A semantic map construction algorithm based on the conditional random field model is studied.The target detection algorithm is used to semantically annotate the key frames,and then the semantic labeling information is fused with the clustering information of the scene segmentation to solve the semantic map.Based on the conditional random field model,three-dimensional semantic annotation is modified under the guidance of the temporal association information of adjacent frames.Experiments show that this algorithm can make up for the rough pointcloud segmentation of small objects in the process of constructing the semantic map of 3D scene.