A FEATURE-BASED DEEP LEARNING APPROACH FOR THE EXTRACTION OF GROUND POINTS FROM 3D POINT CLOUDS

Abstract

Extracting ground points from 3D point clouds is important for sustainable development goals, infrastructure planning, disaster management, and more. However, the irregularity and complexity of the data make it challenging. Deep learning techniques, particularly end-to-end and non-end-to-end approaches, have shown promise for 3D point cloud segmentation and classification, but both require a comprehensive understanding of the features and their relationship to the problem. This paper presents a study on the filtering of 3D LiDAR point clouds into ground and non-ground points using a non-end-to-end deep learning approach. The aim of this research is to investigate the effectiveness of utilizing geometric features and a binary classifier-based deep learning model in accurately classifying point clouds. The publicly available ACT benchmark datasets were employed for training, validation, and testing purposes. The study utilized a k-fold cross-validation method to address the limited availability of training data. The results demonstrated highly satisfactory performance, with validation averages reaching 96.83% for the divided Dataset-1 and an accuracy of 97% for the test set. Furthermore, an independent dataset, Dataset-2, was used to evaluate the generalizability of the trained model, achieving an accuracy of 93%. These findings highlight the potential of the proposed non-end-to-end approach to filtering point cloud data and its applicability in various domains such as DEM and DTM production, city modeling, urban planning, and disaster management. Moreover, this study emphasizes the need for accurate data to achieve sustainable development goals, positioning the proposed approach as a viable option in various studies.