A Deep Learning-Based Method for Overhead Contact System Component Recognition Using Mobile 2D LiDAR.

Area of Science:

• Railway Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Overhead contact systems (OCS) are vital for train power supply, requiring regular inspections to ensure safety.
• Analyzing 2D LiDAR point clouds is a key method for OCS inspection, but recognizing components is challenging due to system complexity.
• Current methods struggle with the accurate semantic segmentation of diverse OCS components from point cloud data.

Purpose of the Study:

• To develop an effective deep learning-based method for semantic segmentation of OCS components from mobile 2D LiDAR point clouds.
• To address the challenges posed by the complex composition of OCS in accurate component recognition.
• To enable both online and batch processing for flexible OCS data analysis.

Main Methods:

• A novel deep learning approach for semantic segmentation of 2D LiDAR point clouds.
• An iterative point partitioning algorithm for enhanced local feature extraction.
• A Spatial Fusion Network module designed for multi-scale local feature learning.
• Scan line by scan line processing for classifying points into object categories.

Main Results:

• The proposed method achieved high accuracy in recognizing sixteen categories of OCS components.
• Mean Intersection-over-Unions (mIoUs) reached 96.12% for online data processing.
• Mean Intersection-over-Unions (mIoUs) reached 97.17% for batch data processing.

Conclusions:

• The deep learning method effectively performs semantic segmentation of OCS components from 2D LiDAR data.
• The approach demonstrates significant potential for improving the efficiency and accuracy of railway infrastructure inspections.
• The developed technique offers a robust solution for recognizing multiple OCS components in complex railway environments.