With the explosive development of intelligent family robot,automatic driving,virtual reality and other emerging industries,the need for semantic analysis and understanding of three-dimensional scene is becoming more and more urgent.At the same time,with the emergence of mature technology of space scanning,three-dimensional sensor technology has made great progress,and 3D data of real scene is more and more easy to obtain.Therefore,three-dimensional semantic analysis is increasingly driven by data.Due to the disorder of point cloud data,the deep learning method based on convolutional neural network cannot directly act on the point cloud,and converting point cloud data into voxel is too expensive to use the three-dimensional convolution method,which is also not suitable for the semantic understanding of indoor scenes.For this reason,a deep learning method based on point cloud three-dimensional data has emerged.PointNet is a pioneer of deep learning based on point cloud.The framework can extract point cloud features effectively,but its performance in semantic analysis is not satisfactory.In this paper,the semantic segmentation of indoor scene point cloud is carried out based on Point Net network framework,and further semantic analysis of indoor scene is conducted,including object detection and instance segmentation in three-dimensional point cloud.The main work and achievements of this paper are as follows:Firstly,this paper introduces an object detection algorithm in indoor scene based on measuring distance.The object detection method based on metric distance in indoor scene proposed in this paper is to obtain the semantic label of each point by the semantic segmentation of indoor scene point cloud,and then use the metric distance defined by us to determine the clustering condition.By the clustering algorithm based on region growth,we implement clustering of target instances in indoor scenes.Compared with the BFS search algorithm in Point Net,we have achieved better results.Secondly,we propose an instance segmentation method based on cosine similarity.we design a method of coupling point cloud feature with instance labels by using cosine similarity,so that under the supervision of instance labels,the characteristic vectors between point clouds belonging to the same instance have a higher cosine similarity,and the cosine similarity between the point clouds that are not part of the same instance is lower.Finally,we improve the Gaussian Blurring Mean-Shift(GBMS)algorithm mean shift algorithm.decouple the feature vectors after cosine similarity coupling to cluster the object in scenes.we tested on S3DIS(Stanford large-scale 3D Indoor Spaces Dataset)dataset and verified the effectiveness of our method.Compared with the related benchmark method,our method achieved better results... |