Multiple-entity Based Heriarchical Classification Method For Airborne LiDAR Point Clouds

With the development of the LiDAR system,and the increasing requirements of industrial applications,the methodology of airborne LiDAR has attracted a large number of advanced approaches.Among these approaches,point cloud classification is the first and most critical step for object recognition,model reconstruction and other further processings.When we take computational primitives into consideration,these point cloud classification methods are able to be divided into three groups,i.e.,individual-point-based ones,segment-based ones,and multiple-entity-based ones.Each group has attracted vast amounts of approaches,however,there are many problems among them.First of all,the individual-point-based methods benefit to the extraction of low-level features,such as edges.However,when they compute features for classification which requires repeated computations and high dimensionality,they enlarge the computational burden.Besides,these methods are easy to be affected by noises in a neighborhood.Although the segment-based methods relieve the computational burden,they rely heavily on their employed point cloud segmentation methods.Up to now,the point cloud segmentation methods mainly extract only one kind of segments,which are difficult to express complex 3D scenes.To deal with this problem,multiple-entity-based methods are proposed,however,they use a hierarchical procedure.The procedure involving several segmentation methods and rather complex steps is defficult to control.Therefore,a multiple-enity-based hierarchical classification method consisting of compact steps is desired.In this thesis,two basic point cloud processing algorithms are proposed firstly,which are utilized to extract 3D edges and express complex 3D scenes,respectively.Then,the two algorithms are combined to design a multiple-entity-based filtering method using iterative graph cuts.Finally,based on the advantages of these proposed methods,and combined with a supervised classification procedure,a multiple-entity-based hierarchical classification method is proposed.The classification method uses three types of entities,and three classification levels.The entities are composed of regular segments,rough segments and individual points.The classification levels are composed of point cloud expression based on entities,ground entities extraction(point cloud filtering),and non-ground entities classification which consists of two sub-levels,i.e.,supervised classification and classification optimization based on semantic rules.In details,point cloud expression based on entities,i.e.,the extraction of three types of entities,is implemented by a step-wise point cloud segmentation method proposed in this thesis.The step-wise point cloud segmentation method clusters the points on the objects with regular geometric structures into regular segments,clusters the points on the objects with irregular shapes into rough segments,and considers the points whose properties are not able to be depicted by their neighborhoods as individual points.Therefore,these entities express the complete properties of 3D scenes.A multiple-entity-based point cloud filtering method using iterative graph cuts is proposed to extract ground entities.First,the method extracts initial ground entities in regular segments.Then,features points of each entity are extracted to replace the whole points in the raw data,which reduces the data size the filter has to process.It is to be noted that the feature points of regular segments are extracted by a 3D edge detection algorithm which is first proposed in this thesis as a basic processing method for point clouds.The edge detection algorithm extracts 3D edges from the raw data without the complements of any image processing and point cloud preprocessing.The non-ground entities classification is implemented by combining the multiple entities and a supervised classification procedure.The method first extracts geometric features for entities.Next,a subset of features is selected using a backward elimination procedure based on the variable importance measure in Random Forests(RF).Then,non-ground entities are classified using the selected features and the trained RF model.At last,a number of semantic rules are defined using the geometric relationships of entities,which are utilized to optimize the initial classification results.The performance of each proposed algorithm is validated by experiments.Based on these experimental results,detailed analysis and discussions are carried out.According to the comparison to the existing methods,the superiorities are validated.In the part of the comprehensive experiments,the open datasets provided by the OpenTopography are employed.Based on the procedure of the multiple-enity-based hierarchical classification method,and the experimental results of each classification level,the utilized algorithms are analyzed to discuss the advantages for the assist of classification.At last,based on the classification results,the accuracies are analyzed and discussed.Besides,the superiorities of the multiple-enity-based heriarchical classification method are validated by comparing with the existing classification strategies.