Point cloud classification using machine learning algorithms and selection of relevant features

Abstract

3D scene classification has become an essential task in photogrammetry, remote sensing, computer vision, and robotics. Point clouds are a data source containing geometric information for 3D world representation. Successful classification results are obtained using the point cloud's geometric information. Machine learning approaches are widely used for point cloud classification. In this study, point cloud classification was performed using random forest (RF) and support vector machine (SVM) algorithms. Geometric features are used to describe each point in the point cloud. However, not every feature may have the same effect on classification. For this reason, the most effective features were determined by applying the filter-based feature selection algorithm. As a result of feature selection, the F1-score value obtained with RF increased by 5.7%, and the F1-score value obtained with SVM increased by 16%.