Research On Dynamic SLAM Method Based On Semantic Segmentation In Indoor Environment

Based on the assumption of static environment,Simultaneous Localization and Mapping(SLAM)has developed well and with good performance in the past few decades,especially the vision-based SLAM algorithm have gained more and more attention owing to the camera's easy access and low cost.However,the existence of moving objects in real scenes greatly limits the application of visual SLAM in dynamic environments.How to deal with moving objects in dynamic environment has become the key to improve the performance of SLAM systems.This thesis designs a visual SLAM system for dynamic environments,constructs a three-dimensional dense map containing only static information,and improves the accuracy and robustness of the SLAM system in dynamic scenarios.The specific research contents are as follows:Firstly,the thesis introduces the Mask R-CNN,an instance segmentation network,and tests Mask R-CNN on the datasets and real scenes.On this basis,two lightweight attention modules,position attention module and channel attention module,are introduced to Mask R-CNN aiming at the misclassification and inaccurate segmentation problems of CNN(Convolutional Neural Network,CNN).By considering the interdependence between each position and each channel in the image,the global information is aggregated,the global dependence and discrimination ability of the network is enhanced,and the classification accuracy is improved.The experimental results show that the Mask R-CNN network with the attention mechanism has a smaller loss value and faster convergence,which alleviates the problem of inaccurate segmentation.Secondly,in view of the problem that Mask R-CNN cannot distinguish the motion attributes of objects without prior information,the thesis uses a dense pyramid optical flow method to detect moving objects,and designs an image feature point extraction method to reduce time cost of constructing dense optical flow field.Combining the dynamic information of the objects in the scene detected by the optical flow method and the semantic information of the objects output by Mask-RCNN,the moving objects in the scene are detected and segmented,and the mask of the moving objects is removed from the image.Experimental results show that the algorithm can accurately detect moving objects and eliminate them from the scene.Then,removing dynamic objects will also remove the background area covered by them.This thesis believes that the static information contained in the background area is useful for map construction and reuse.To solve this problem,it designs a background repair network based on deep learning.For the blank area left by culling dynamic objects in the current frame,static information from other views is used to fill the background occluded by dynamic objects,and experiments and analysis of the background repair network on the dataset verify the effectiveness and feasibility of the algorithm.Finally,in view of the low efficiency of the point cloud map and the large consumption of memory resources,this thesis uses an algorithm based on the Surfel model to reconstruct the three-dimensional map,which improves the reconstruction efficiency.The experimental results show that the map established by the algorithm effectively removes the dynamic objects in the environment and inpaints the static background,which solves the problems of object ghosting in the map and low camera pose estimation accuracy.