Research On Registration Method Of 3D Point Cloud And Floor Plan Based On Feature Matching

For most of the existing buildings,there are often corresponding building floor plans,and the registration of 3D point clouds and floor plans is a valuable and niche research field that needs to be explored urgently.Studying the registration method of 3D point cloud and floor plan will help to solve many industry problems,and bring new ideas to the research of surveying and mapping exploration,indoor navigation,3D reconstruction,BIM system,SLAM,intelligent robot and other fields.Solving such problems has High academic value and practical application value.3D point cloud and floor plan registration problem is closely related to the of the problem of robot localization on the floor plan.Many scholars have proposed Monte Carlo Localization method to solve such problems.However,the Monte Carlo positioning method has particle degradation and slow convergence problem.It's not suitable for dealing with large-scale scenes.Combining the latest deep learning theories and methods and inspired by point cloud registration,this thesis proposes 3D point cloud and floor plan registration method based on feature matching and it realizes the end-to-end highprecision registration of the both.The main research work of this thesis is as follows:1)3D point cloud and floor plan feature extraction method.Firstly,a preprocessing method of 3D point cloud and floor plan is proposed to convert the original data is converted into a data form usable by the model proposed in this thesis.Secondly,an improved dynamic graph convolutional neural network method is proposed to extract the cross-dimensional features of the both.For the problem that the point coordinates can provide less information,add additional point cloud features including normal vector and curvature information to the network input;Aiming at the problems of low utilization of geometric information of similar features in Edge Conv and degeneration of edge features in deep networks,this thesis proposes an improved method of introducing edge features into geometric structure features;Aiming at the problem of gradient disappearance in Edge Conv deep network,this thesis proposes a residual network structure;In order to realize the feature information exchange between 3D point cloud and floor plan,a feature fusion method based on graph attention network is proposed.2)3D point cloud and building plan feature matching method.On the basis of obtaining the feature description,in order to better separate the point pairs without corresponding relationship,establish matching constraints and optimize the matching results,a Sinkhorn optimal transmission algorithm is proposed,and augmented points are introduced to achieve matching optimization,combined with singular value decomposition.For the point cloud registration method,the first 3D point cloud and floor plan registration network model FPST-Net v1 is proposed in this thesis,and these method performance is verified by experiments.3)End-to-end registration network of 3D point cloud and floor plan.In order to improve the accuracy of the registration results,this thesis introduces a differentiable alignment method,and proposes a translation correction network to correct the translation error in the z-axis direction of the registration results,and at the same time improves the loss function,and finally proposes an improved FPST-Net v2.This network model can realize end-to-end training and prediction,and finally verify the actual performance of the method proposed in this thesis through experiments.Experiments on CVC-FP,ROBIN-FP and TUM Indoor datasets show that the proposed method can effectively achieve the registration of 3D point cloud and floor plan,and shows high accuracy,strong generalization and anti-interference performance.