Research On Semantic Map Construction Algorithm Of Indoor Scene Based On Deep Learning

The service robots based on semantic map navigation are capable to execute commands in a natural interactive way between individuals through the semantic information in the environment,which has a widespread application prospect.However,because the real indoor environment is an intricate environment combining dynamic and static,the current vision Simultaneous Localization and Mapping(SLAM)technology remains the problem of insufficient robustness when applied in complex indoor environments,which affects the construction of semantic maps.To address the issues mentioned above,the dissertation conducted the research as following:1)An improved DeeplabV3+ semantic segmentation network model is proposed,which improves the boundary blur and breakage of indoor target overlap and small regions in the semantic segmentation algorithm.By replacing the shallow feature extraction network,the complexity of the model is reduced,the running speed of the model is accelerated,and the decoder is improved by multi-layer information fusion to restore the detailed information of the target boundary.In order to further refine the target boundary,based on the principle of feature sharing,a multi-scale information boundary extraction network for predicting the target boundary is designed.The experimental results demonstrate that compared with the DeeplabV3+ model of the Res Net-50 network,the segmentation accuracy of the improved model is increased by 2.5%,and the boundary contour of the target is more obvious.2)A loop detection method to remove dynamic targets is proposed,which relieves the interference caused by the diverse motion of indoor objects on the loop detection method.Firstly,the environment dynamic target library is established.Then,combine the L-K optical flow method and the improved DeeplabV3+ semantic segmentation network model to detect dynamic targets,and remove all feature points on the target;Further,use historical key frames to compensate for feature points in the target removal area;Finally,based on the bag-ofwords model,and adopt the ORB-SLAM2 system’s loopback verification strategy for loopback detection.The experimental results manifest that in complex indoor environment,compared with the loop detection algorithm in ORB-SLAM2,the loop accuracy of the algorithm in this thesis is improved by 15.2% on average,and it possesses higher robustness in various dynamic scenarios.3)Combined with the improved semantic segmentation and loop detection methods,a visual SLAM system for building indoor semantic maps is built on the basis of the ORBSLAM2 framework.The experimental results illustrate that compared with the ORB-SLAM2 system,the visual SLAM system established in this thesis has higher robustness in complex indoor environment and is capable of building visual semantic maps online.