Machine Learning-Based Supervised Classification of Point Clouds Using Multiscale Geometric Features

Abstract

3D scene classification has become an important research field in photogrammetry, remote sensing, computer vision and robotics with the widespread usage of 3D point clouds. Point cloud classification, called semantic labeling, semantic segmentation, or semantic classification of point clouds is a challenging topic. Machine learning, on the other hand, is a powerful mathematical tool used to classify 3D point clouds whose content can be significantly complex. In this study, the classification performance of different machine learning algorithms in multiple scales was evaluated. The feature spaces of the points in the point cloud were created using the geometric features generated based on the eigenvalues of the covariance matrix. Eight supervised classification algorithms were tested in four different areas from three datasets (the Dublin City dataset, Vaihingen dataset and Oakland3D dataset). The algorithms were evaluated in terms of overall accuracy, precision, recall, F1 score and process time. The best overall results were obtained for four test areas with different algorithms. Dublin City Area 1 was obtained with Random Forest as 93.12%, Dublin City Area 2 was obtained with a Multilayer Perceptron algorithm as 92.78%, Vaihingen was obtained as 79.71% with Support Vector Machines and Oakland3D with Linear Discriminant Analysis as 97.30%.

A Two-Step Feature Extraction Algorithm: Application to Deep Learning for Point Cloud Classification

“…This open access data set was collected in Vaihingen, Germany using a Leica ALS50 system with a scanning height of 500m having a field of view 45 o . It has been frequently used for per point labelling (Blomley et al, 2016;Atik et al, 2021). This data set has point densities of 4-6/m 2 .…”

“…Some DL algorithms for classification are considered end-to-end as they use point coordinates, normalized coordinates and/or a few features such as intensity and colors (Qi et al, 2017a, b;Thomas et al, 2019;Hu et al, 2020), but many are not (Hsu and Zhuang, 2020;Nurunnabi et al, 2021a) and rely upon using hand-crafted features such as point normal and curvatures as inputs, instead of just points. Many of the latter feature-based DL algorithms use multi-scale (Thomas et al, 2018;Cabo et al, 2019;Atik et al, 2021) and/or multi-type (Blomely et al, 2016;Weinmann and Weinmann, 2019) features to improve classification performance. Laser scanning based point clouds are challenging to classify as they are usually unstructured, having highly variable point density and irregular data format, and are typically capturing sharp features (e.g., edges and corners) and arbitrary surface shapes.…”

Graph Attention Feature Fusion Network for ALS Point Cloud Classification

“…This kind of method is highly dependent on the threshold, and therefore has poor adaptability. The supervised method provides hand-engineered features to traditional machine learning algorithms [ 9 , 10 , 11 , 12 ] for classification. This method does not have the ability to learn high-level features; in fact, it is difficult to further improve the classification accuracy.…”

“…Point cloud classification usually involves assigning a category label to each point, which is often referred to as point cloud semantic segmentation in computer vision. In early studies, researchers classified point clouds by employing hand-engineered features and traditional classifiers [ 9 , 10 , 11 , 12 ] or preprocessed the point clouds before classification [ 13 , 14 ]. These methods belong to traditional machine learning methods, which fail to learn high-level features, whereas the methods based on deep learning techniques can further improve the classification accuracy due to the ability to learn high-level features.…”

Automatic Classification of Urban Objects From Mobile Laser Scanning Data

“…After attributes are extracted from the point cloud, a class label must be assigned to each point. For this purpose, machine learning methods (e.g., Weinmann et al, 2015;Xiao et al, 2016;Zheng et al, 2017;Sun et al, 2018;Thomas et al, 2018;Wang et al, 2018;Atik et al, 2021;Seyfeli and Ok, 2022a) or deep learning methods (e.g., Balado et al, 2019;Guo and Feng, 2020) can be utilized.…”