Research On Semantic Segmentation Algorithm Of 3D Point Cloud Based On Deep Learning

In the field of computer vision,scene understanding based on 3D point cloud has been favored by most researchers.Among them,3D point cloud semantic segmentation technology is widely used in indoor scene analysis,robot environment perception,automatic driving and other fields.The early 3D point cloud semantic segmentation method mainly uses mathematical calculations to divide the scene into multiple non-overlapping regions.However,the accuracy of this method is easily affected by noise and scene complexity,and lacks robustness.Recently,with the introduction of the Point Net,the end-to-end processing of 3D point clouds based on deep learning technology has gradually become the mainstream,but there are still some deficiencies.The main problem is that the consideration of the segmentation effect of the edge area is ignored,and the effective extraction of point cloud features cannot be guaranteed.To address the above problems,this thesis explores the research on semantic segmentation and feature extraction of 3D point cloud based on deep learning,with the main work and contributions as follows:(1)To address the problem of ineffective segmentation of edges,this thesis proposes a 3D point cloud semantic segmentation network model based on semantic search and superpoint attention.Firstly,semantic search is used to define the center coordinates and features of superpoints,ensuring semantic consistency of points within superpoints.Secondly,the attention mechanism is used during the superpoint segmentation stage to extract features of different levels and achieve semantic segmentation of superpoints.Then,the segmentation results are applied to each point through a label feedback module.Finally,experimental results demonstrate that the semantic segmentation results of this method are superior to other point cloud segmentation methods,and the network model also has good robustness.(2)To address the problem of insufficient key feature extraction,this thesis proposes a plug-in based feature extraction module that integrates multiscale information.Firstly,this module is applicable to scenes with different resolutions,and can transform input point cloud data into a high-dimensional feature sequence.Secondly,this thesis uses self-attention mechanisms in channel fusion to focus on the correlation between dimension channels,and extract point feature information from the channel perspective.Then,this thesis analyzes the feature similarity in spatial fusion to obtain the geometric structure information and contextual dependence information of the point cloud.Finally,the effectiveness and reusability of this module are demonstrated,and it is also proven that the network model based on this module has a good understanding of 3D scenes.