Bilinear Distance Feature Network for Semantic Segmentation in PowerLine Corridor Point Clouds.

Area of Science:

• Computer Vision
• Machine Learning
• Geospatial Analysis

Background:

• Semantic segmentation of power transmission line corridor point clouds is vital for detecting tree barriers.
• Challenges include massive data, disordered distribution, and non-uniformity, hindering feature extraction.
• Existing methods often neglect spatial information, limiting geometric shape understanding.

Purpose of the Study:

• To enhance deep expression of spatial geometric information in segmentation networks.
• To propose BDF-Net, an improved method based on RandLA-Net for powerline corridor point cloud segmentation.
• To improve the accuracy and efficiency of powerline tree barrier detection.

Main Methods:

• Developed BDF-Net incorporating Spatial Information Encoding, Bilinear Pooling, and Global Feature Extraction blocks.
• Spatial Information Encoding captures local structure using relative coordinates and distances.
• Bilinear Pooling integrates point cloud features with spatial geometric representations.
• Global Feature Extraction enhances semantic understanding using point position ratios.

Main Results:

• BDF-Net achieved significant performance improvements on the PPCD dataset.
• Achieved an Overall Accuracy (OA) of 97.16%, mean Intersection over Union (mIoU) of 77.48%, and mean Accuracy (mAcc) of 87.6%.
• Outperformed RandLA-Net by 3.03% (OA), 16.23% (mIoU), and 18.44% (mAcc), demonstrating superiority over state-of-the-art methods.

Conclusions:

• BDF-Net effectively enhances spatial geometric feature extraction for semantic segmentation.
• The proposed method significantly improves point cloud segmentation in power transmission line corridors.
• BDF-Net offers a superior solution for powerline tree barrier detection and related geospatial analysis tasks.