Content-Sensitive Multilevel Point Cluster Construction for ALS Point Cloud Classification

Abstract

Airborne laser scanning (ALS) point cloud classification is a challenge due to factors including complex scene structure, various densities, surface morphology, and the number of ground objects. A point cloud classification method is presented in this paper, based on content-sensitive multilevel objects (point clusters) in consideration of the density distribution of ground objects. The space projection method is first used to convert the three-dimensional point cloud into a two-dimensional (2D) image. The image is then mapped to the 2D manifold space, and restricted centroidal Voronoi tessellation is built for initial segmentation of content-sensitive point clusters. Thus, the segmentation results take the entity content (density distribution) into account, and the initial classification unit is adapted to the density of ground objects. The normalized cut is then used to segment the initial point clusters to construct content-sensitive multilevel point clusters. Following this, the point-based hierarchical features of each point cluster are extracted, and the multilevel point-cluster feature is constructed by sparse coding and latent Dirichlet allocation models. Finally, the hierarchical classification framework is created based on multilevel point-cluster features, and the AdaBoost classifiers in each level are trained. The recognition results of different levels are combined to effectively improve the classification accuracy of the ALS point cloud in the test process. Two scenes are used to experimentally test the method, and it is compared with three other state-of-the-art techniques.

Higher-Order Conditional Random Fields-Based 3D Semantic Labeling of Airborne Laser-Scanning Point Clouds

“…Multilevel clustering can better explore the contextual information between the objects in outdoor scenes. Although there exists a multilevel cluster construction algorithm in [13], the size of generated clusters is linearly increased, causing no apparent feature discrepancy of the objects. (c) Higher-Order CRF Optimization: To optimize the point labeling using the proposed higher-order CRF model, we not only consider the adjacent clusters but a wider local area based on the neighborhood relationship between the clusters.…”

“…Accuracy is the most intuitive performance measure and is the ratio of correctly predicted observations to the total observations, i.e., accuracy = (TP + TN)/(TP + FP + TN + FN) [50]. F 1 -score is the weighted average of the precision and recall and is defined as: F 1 -score = 2 × (recall × precision)/(recall + precision) [13]. Therefore, this score takes both false positives and false negatives into account.…”

“…However, in single point-based methods, the representation of the geometric structure and contextual information in individual point clouds has not been fully used, resulting in lower accuracy of point labeling. To solve this problem, cluster-based classification approaches [12,13] have been proposed by adding a segmentation step prior to classification. In this case, the basic processing units are not individual point clouds but the segmented clusters, from which the high-level discriminative features are derived and used for the final classification.…”

“…In this case, the basic processing units are not individual point clouds but the segmented clusters, from which the high-level discriminative features are derived and used for the final classification. For example, Xu et al [13] proposed a cluster-based ALS point cloud classification method. This method constructs content-sensitive multilevel point clusters based on the density distribution of ground objects.…”

“…These multilevel cluster's features are finally used for classification. The advantage of pre-segmentation in works [12,13] is that it is not affected by discontinuities and the poor quality of the input point clouds because the contiguous point clouds are grouped into larger entities that share similar properties or relationships. This results in homogeneous regions that are easier to interpret.…”