Research On Semantic Segmentation Of Indoor Building Structure Point Cloud Based On Deep Learnin

As the core method of indoor scene understanding,point cloud semantic segmentation of indoor building structure has become a research hotspot in recent years.Precise semantic segmentation of point cloud is a key technology in indoor navigation,3-dimensional reconstruction,robot inspection and other domains,which has great significance both in science and application.The current networks pay too much attention to the extraction of local features,yet neglects the influence of multi-scale information,global information and long-range dependence.This paper conducts an in-depth study on the relationship between local features and multi-scale information,the relationship between local features and global features,and long-range dependence in the semantic segmentation of point cloud for indoor building structures.An end-to-end point cloud semantic segmentation network CGIBS-Net for indoor building structures is proposed,including the following three parts:(1)Aiming at the problem that feature encoding of edge points usually contains multiple categories of semantic information,leading to inaccurate object contour segmentation,a local awareness branch(LPB)is proposed and added into the multi-scale context fusion module,which can better learn and sense the local structure of point cloud in the multi-scale environment.Thus,the local characteristics of point cloud can be perceived more fine-grained.(2)Aiming at the problem that insufficient attention was paid to global information while optimizing the local coding structure of multi-concern cloud,a double-channel cross-grouped self-attention mechanism(CGSA)is proposed to comprehensively consider local and global feature extraction.The feature and position-direction information of point cloud can be processed simultaneously by CGSA,and all the information is selectively enhanced.(3)In order to solve the problem of inaccurate prediction caused by the spatial distance between two objects of the same semantic category,a self-learning weight spacechannel attention mechanism fusion(S-C)module is designed in the decoder.S-C is an effective way to capture long range dependencies,similar features are ensured to be related with each other by two complementary attention modules no matter how far apart they are.Compared with the Point Net++ model,the overall accuracy o Acc,the average class accuracy m Acc and the average class union m Io U of the model is increased by 6.9%,9.4%and 7.8% respectively on the S3 DIS dataset by using the 6-fold cross validation.Compared with the CSANet model,the o Acc,m Acc and m Io U in the S3 DIS dataset of this model are increased by 3.5%,2.4% and 5.1%,respectively.