|
Evaluation of RandLA-Net Model for Semantic Segmentation of 3D LiDARHD Point Clouds in Complex Urban Environments through Preprocessing Optimization
|
Javad Sadidi *   , Hamzeh Rafizadeh , Hani Rezaian |
| Kharazmi University |
|
|
Abstract: (11 Views) |
Semantic segmentation of 3D point clouds is a key challenge in LiDAR data processing and the applications which are related to computer vision, robotics, and urban mapping. In this study, the performance of the RandLA-Net deep learning model for semantic segmentation of 3D point clouds is investigated. The dataset used in this study is the Dutch LiDARHD data called AHN4, which contains accurate elevation information from different areas, especially the urban and natural environments. To improve the model's accuracy, a data preprocessing pipeline including denoising, normalization, and data augmentation was applied. To assess the model's performance, the Intersection over Union (IoU) metric was used, which measures the percentage of overlap between the predicted and ground truth segments for each class and serves as a precise criterion for evaluating segmentation quality. The results indicated that the RandLA-Net model achieved high performance in segmenting classes such as ground (92/3% IoU), vegetation (93/6% IoU), and buildings (93/7% IoU). However, the complex and underrepresented classes like bridges (65/4% IoU) were identified with lower accuracy. A comparison of the implemented model with baseline methods such as PointNet++ and KPConv shows that RandLA-Net provides higher accuracy with lower computational cost. The overall results of this study, with an overall average community-to-community ratio (mIoU) of 87/1% on the LiDARHD dataset, confirm the effectiveness of deep learning in increasing the accuracy and efficiency of the semantic segmentation of 3D point clouds for high-frequency classes. At the same time, these findings highlight the ongoing challenges in correct separation of the sparse classes in complex urban environments.
|
|
| Keywords: 3D Point Clouds, Semantic Segmentation, Deep Learning, RandLA-Net, LiDAR, Preprocessing |
|
|
Full-Text [PDF 2371 kb]
(6 Downloads)
|
Type of Study: Research |
Subject:
GIS
Received: 2025/05/2 | Accepted: 2025/09/3 | ePublished ahead of print: 2026/01/14 | Published: 2026/01/14
|
|
|
|
|
|
|
| Send email to the article author |
|
|
