Research On Semantic Segmentation Method Of Indoor Point Cloud Based On Deep Learning

In recent years,the indoor 3D model construction technology progresses and develops constantly and has been widely applied to indoor navigation,BIM model construction,and virtual reality(VR)domains.3D laser point cloud is the most widely used data source for indoor 3D model construction and 3D laser point cloud segmentation one of the key technologies of constructing indoor 3D models.Since the widespread application of deep learning in the he field of computer vision,it is directly applied to 3D points,deep-learning-based semantic segmentation method for3 D point cloud has attracted more and more attention and researching.PointNet,the very first deep neural network that directly input point cloud,has a significant meaning.However,this network mainly extracts the global features of point cloud and neglects the learning and extraction of local features.Aiming at the shortcomings of the original PointNet network,this article improve that and put forward PointNet-Pro bases on PointNet and construct an indoor data set for comparative experiment to validate the effectiveness of PointNet-Pro.The details are as follow:(1)According to the intensity gap between different architectures,put the information of 3D coordinate,color and intensity into the feature space bases on the characters that point cloud is disordered and has no specific spatial topology.In the process of constructing the network,this article used a feature-extracting algorithm base on R-nearest local area.Use different radius of sphere as threshold to search the local area for each point and extract the high dimension features from the local area as the local feathers of input data.In order to enhance the ability of the network to extract features,replace the 3 × 1 convolution kernel that the original PointNet network used to a larger 3 × 3 convolution kernel.Accroding to the semantic segmentation experiment using the public S3 DIS data set from Stanford and the comparison with PointNet segmentation result,the modified network improves the precision of every categories and the IoU gets 4.1% raised.(2)Construct a dense indoor point cloud contains various indoor scenes and 8different categories.In the process of creating the data set,firstly use a station-mounted scanner to scan the main indoor scene of school's architecture and retain the reflection intensity information during the scanning process on purpose.After getting the original point clouds,preprocess them for registering,disnoising,thinning and other operations.After the preprocessing,label every pointof the point cloud and divide the data sat as training set,validation set and testing set.Finally,the result of comparative experiment using the data set shows the modified network get the precision that the accuracy of walls and doors higher are than 80% and the average segmentation accuracy of all architectural elements reach66%.