Indoor Scene Point Cloud Perception And 3D Geometric Expressio

With the emergence of concepts such as smart cities and digital twins,there has been an increasing focus on the foundational infrastructure of urban geospatial information,especially indoor 3D models.Due to the rapid development of various 3D point cloud technologies based on laser radar sensors,this form of data can effectively depict complex real-world scenes and provide valuable data support for 3D modeling.However,3D point clouds are unstructured and do not contain semantic information.Currently,modeling based on 3D point clouds generally requires a combination of manual feature extraction methods and prior knowledge to extract the structural and semantic information of the 3D point clouds under the strong Manhattan assumption.To address these limitations,this paper improves the semantic categorization of indoor point cloud datasets and proposes a method that reduces the dimensionality of 3D point clouds to a 2D space for processing and completing 3D indoor modeling.This method combines deep learning-based point cloud semantic segmentation techniques with graph-cut optimization algorithms.The main contents of this paper are as follows:(1)In order to improve the accuracy of point cloud semantic segmentation and facilitate reconstruction work,this paper adopts two methods: improving the organization of indoor point cloud datasets and improving the deep learning point cloud semantic segmentation network.Firstly,in the indoor point cloud dataset,movable features are merged into clutter categories,and the topological organization of the dataset is scenarized.The 13 semantic categories of the S3 DIS dataset are changed to 6 to reduce the complexity of semantic categories and improve the average intersection-over-union of the KPConv deep learning method by about 5.6%.Secondly,this paper designs a deep learning point cloud semantic segmentation network,which proposes an explicit spatial encoding and channel attention fusion mechanism module on the deep learning network structure of the autoencoder.The explicit spatial encoding can alleviate the problem of difficult learning of small components under the scene level by aggregating the position information,relative position relationship,and distance information of neighboring points with the central point;the channel attention fusion mechanism fuses and weights the explicit local spatial encoding with the high-dimensional features learned by the network,thereby helping the network to recognize the missing windows and doors of the point cloud.On the reprocessed dataset,the average intersection-over-union of the network in this paper reached 78.5%,which is about 2.4%higher than the segmentation results of KPConv.(2)This paper proposes a geometric modeling method for indoor room-scale environments.Firstly,the entire point cloud is directly projected to generate an indoor semantic map,while the semantic point cloud of walls and windows is projected to generate a wall semantic map.Secondly,the two semantic maps are subtracted,and morphological segmentation is applied to obtain a room segmentation map.Then,the wall semantic point cloud is horizontally sliced,and a simplified linear primitive is extracted using the random sampling consensus algorithm and mean shift algorithm,resulting in an over-segmentation polygonal unit for indoor space.The graph cut algorithm is applied to the polygonal units,and based on the semantic map,data terms and smoothness terms are designed to complete the segmentation of indoor and outdoor polygonal units.By overlaying the room segmentation map with the polygonal units,polygonal units with semantic information at the room scale are obtained.Finally,the rooms are vertically extended to complete the geometric modeling,and the accuracy of the indoor room segmentation and reconstruction algorithm is verified on point clouds from three different indoor areas.