| The Chinese culture has a long history,and a large number of cultural treasures are left behind from each dynasty to each generation.Among them,ancient buildings are important treasures of Chinese culture and the core of inheriting Chinese culture.The information of early ancient buildings is mainly stored in the form of text,drawing,taking pictures,etc.With the maturity of 3D reconstruction technology,a large amount of surface information of ancient buildings is preserved in the form of 3D point clouds.The large wooden structure of ancient Chinese buildings has its own characteristics,and each component is the basic unit of the main body of ancient buildings.Classifying the ancient building components from the three-dimensional point cloud is a crucial step in realising the efficient modelling of large wooden structures.Classification of point cloud data mainly includes traditional classification algorithm and machine learning classification algorithm.The algorithms mainly include nearest neighbour classification,Cluster-based,support vector machine,random forest and so on.Although traditional classification algorithms are constantly improving to improve the classification effect,most of these algorithms are aimed at small scene tasks,which are challenging to classify massive point clouds of ancient buildings.Most algorithms need to set feature descriptors,a large number of thresholds and parameters artificially,and the steps are cumbersome.Not intelligent enough.In recent years,the rapid development of deep learning provides a new research idea for the classification of point cloud components in ancient buildings.Given the above analysis,this paper uses the classical depth neural network PointNet,which directly takes the point cloud as the input,to classify and extract the columns and foreheads of the ancient buildings of the Forbidden City.Point cloud classification based on deep learning reduces the degree of manual intervention and makes the process of component classification more intelligent.Based on the classification,this paper uses the method based on statistical information to classify the columns and weights extracted from deep learning more finely,which improves the accuracy of classification results.In the point cloud classification of ancient buildings based on deep learning,this paper improves the PointNet network to improve the classification effect.The main contents of the improvement are as follows:(1)analysing the characteristics of point cloud input neural network,adding colour features,reflection intensity and other features into the data set of self-built ancient buildings;(2)improving the local feature extraction ability,selecting seed points from the input data.The scale region is divided with the seed point as the center,and the points in the region are taken as the neighbourhood points of the seed point,and the local features in the region are extracted,and then the local features are connected with the global features.Through the above improvements,we can improve the ability to learn the features of the whole network.In the aspect of constructing fine classification based on statistical information,considering the geometric characteristics of the standard components of ancient buildings and the differences in sizes among different components,this paper takes the standard models of columns and pillars as a reference,applies the D1 shape distribution function based on information statistics and matches the components extracted by the deep learning method with the standard models to determine the specific categories of components,so as to achieve fine classification.Because the simple shape distribution function only considers the distance feature,the obtained eigenvalues can not fully characterise the model.In this paper,the method in which the area is used as the weight factor is applied to resample the model and calculate the eigenvalues in order to obtain more stable similarity values and improve the matching accuracy.Because of the above research results,a three-dimensional experimental platform based on visual studio2013+QT5+OSG is developed in this paper to realise the functions of model display,sampling visualisation,matching,etcThe research results of this paper improve the accuracy of point cloud classification of ancient buildings,provide reference data for studying the classification,reconstruction and health monitoring of large wooden structures of ancient buildings,and have specific practical significance for the scientific and technological protection of ancient buildings. |
