Research On 3D Point Cloud Data Segmentation And Classification Algorithm For Outdoor Scenes

In computer science,computer vision is an important research field,and contains a series of sub-problems.As a widely-studied sub-problem,3D data processing has become a hot research topic,and has been developed in various fields such as reverse technology,3D modeling,aerial surveying,machine vision,human-machine dialogue and virtual simulation.In digital cities,intelligent transportation and other comprehensive platforms,outdoor scene 3D point cloud data has become an indispensable data source for smart cities and a key technology for urban virtual reality.For outdoor scene,it is comprehensive and complex,and owns a variety of features.However,when collecting outdoor scene data,a large number of interference and noise is often intruded,which would increase the difficulty and cost of data applications.To solve these problems,we can use several available techniques in 3D point cloud data processing,such as segmentation and classification.These two techniques can semanticize point cloud data and reduce the application difficulty of point cloud data.Therefore,research and application on cloud data segmentation and classification for outdoor sites are widely recognized today.This thesis first conducts a comprehensive investigation on point cloud segmentation and classification algorithm at home and abroad.Then the purpose of this study is proposed: provide a solution containing applicable combination of segmentation algorithm and classification algorithm for 3D outdoor scene point cloud,through an integrated study,which includes analyzing,evaluating and comparing of different types of algorithms.In the end,an outdoor scene cloud classification algorithm combined with progressive morphological filtering and semi-supervised support vector machines is proposed.The main contributions are listed as follows:This thesis studies the classic point cloud data segmentation methods at home and abroad including random sampling consensus based segmentation,region growing based segmentation and cluster based segmentation.In terms of cluster based segmentation methods,it includes K-means clustering and supervoxel clustering.The basic principles,advantages and disadvantages are analyzed in each algorithm respectively.The proposed segmentation solution is based on a progressive morphological segmentation algorithm for ground points removing.By leveraging the active structuring element,the size of structuring element is continuously increased,ground and non-ground points is identified accurately.For the classification of non-ground data points,this paper evaluates two types of classification algorithms including iterative self-organizing data analysis technique and conditional random field.Based on the dimension characteristics of non-ground points,we develop an effective and efficient classification model,which uses the semi-supervised support vector machine as the base model.In this paper,the outdoor scene cloud for verifying the effectiveness of the algorithm is collected by the International Photogrammetry and Remote Sensing Society.Compared with the traditional point cloud segmentation classification method,the proposed method has the characteristics of low data acquisition cost and high computational accuracy.Such strong points broaden the prospect of the proposed method in practical applications.