Classification of Point Cloud Data in Road Scenes Based on PointNet+.

Area of Science:

• Geoinformatics
• Computer Vision
• Artificial Intelligence

Background:

• Point cloud data is crucial for urban road infrastructure management.
• Manual classification of point cloud data is time-consuming and error-prone.
• Deep learning offers potential for automated feature extraction and classification.

Purpose of the Study:

• To develop an automated deep learning method for point cloud classification in road scenes.
• To enhance the PointNet++ framework for improved accuracy and robustness.
• To create a comprehensive dataset for training and evaluating road scene point cloud models.

Main Methods:

• Utilized Princeton ModelNet40, ShapeNet, and Sydney Urban Objects datasets.
• Implemented Farthest Point Sampling (FPS), random transformations, and Gaussian noise injection for data augmentation.
• Integrated a point-filling method into the PointNet++ preprocessing module.
• Employed Multi-Scale Grouping (MSG) and Single-Scale Grouping (SSG) schemes for model training and prediction.

Main Results:

• Achieved an average training accuracy of 86.26% with peak single-instance accuracy of 98.54%.
• Reached a test set accuracy of 97.41% with a category accuracy of 84.50%.
• Demonstrated high-precision object recognition in complex road environments.

Conclusions:

• The developed deep learning model successfully classifies road scene point clouds.
• The study provides valuable insights into point cloud data processing for urban infrastructure.
• Automated classification enhances the efficiency and precision of road surveying and management.