An aerial point cloud classification using point transformer via multi-feature fusion

Area of Science:

• Computer Vision
• Machine Learning
• Geospatial Data Analysis

Background:

• Point Transformer models excel at capturing local and global context in point cloud data.
• Existing methods struggle with instance structure preservation in large-scale aerial point clouds, hindering fine-grained classification.
• Loss of detail in point cloud tokenization limits feature representation for complex urban environments.

Purpose of the Study:

• To propose a novel Point Transformer-based Multi-feature Fusion (PTMF) Network for enhanced aerial point cloud classification.
• To address limitations in instance structure representation and fine-grained object classification.
• To integrate geometric features to complement existing contextual feature extraction methods.

Main Methods:

• Developed a PTMF Network integrating geometric features into the Point Transformer architecture.
• Employed a multi-stage fusion of down-sampled geometric and inherent point cloud features.
• Utilized a Transition Up module for effective up-sampling of mapping features.

Main Results:

• Achieved significant improvements in classification accuracy on the SensatUrban and DALES datasets.
• Attained mean Intersection over Union (mIoU) scores of 63.52% on SensatUrban and 82.18% on DALES.
• Demonstrated superior performance compared to existing state-of-the-art methods.

Conclusions:

• The PTMF Network effectively enhances feature representation for large-scale aerial point clouds.
• Explicit integration of geometric features significantly improves fine-grained object classification.
• The proposed method offers a robust solution for analyzing complex urban aerial point cloud data.