A Study Of SLAM In Dynamic Scenes And Semantic Mapping Based On Semantic Segmentation

With the continuous promotion of service robots and unmanned driving,Simultaneous Localization and Mapping(SLAM)has become one of the most popular research topics in recent years.This paper studies the problems of poor localization effect and low map quality of visual SLAM in comlpex dynamic scenes,and proposes an RGB-D SLAM system that can remove interference from moving objects and supports dense semantic mapping.This paper proposes a feature point selection algorithm in dynamic scenes,including a dynamic feature point detection process based on epipolar geometry and semantic segmentation,and a stable feature point selection process based on information theory and semantic segmentation.Candidate dynamic feature points are obtained through optical flow and epipolar constraints and examined with semantic class and classification entropy,solving problems such as missed detection,misrecognition,and geometric constraint failure under extreme conditions in dynamic feature point detection process,and improving the accuracy of camera pose estimation.By calculating the mutual information of the feature point and the camera pose while introducing the classification entropy to measure the stability of the feature points,saving the feature points that make the pose with larger entropy reduction and has smaller classification entropy at the same time,the local map can be improved to improve the pose estimation accuracy,and map scale is reduced whilie the ability of map for long-term localization is improved.This paper improves the octree map and designs a dense semantic mapping method on this basis.To speed up the map reconstruction,this paper adds a sub-map mechanism to replace the reorganization of a large number of key frames with the fusion of fewer sub-maps.At the same time,a global optimization mechanism is added.The initial frame of the sub-map is used as a representative of the sub-map to participate in global optimization.To ensure the full map is highly consistent with the real environment,when the sub-maps are fused,each sub-map is converted to the global SLAM coordinate system according to the pose of the initial frame after loop optimization.In order to expand the semantic layer of the map,semantic classes of pixels are introduced in the process of generating point clouds from depth maps,and the dynamic voxel grids are removed and the semantic classes are passed on.Finally,different semantic observations in the same grid can be concluded based on the update strategy of semantic class hit times.This paper comprehensively evaluates the performance of the system through experiments such as dynamic feature point detection,camera pose estimation,and dense map construction.This experiment selects several sequences with different numbers of moving objects from RGB-D data set,and compares the results with the classic visual SLAM open source solution.The experimental results prove that the feature point selection algorithm designed in this paper can deal well with various complex dynamic scenarios.In high dynamic scenes,the accuracy of pose estimation is improved by about 90%.The experimental results also prove that the dense semantic mapping method designed in this paper is suitable for dynamic scenes.The map information is more abundant than the feature point map,and the reconstruction time efficiency is improved about 80% compared with the original octree map.