Enhancing rail safety through real-time defect detection: A novel lightweight network approach

Area of Science:

• Computer Vision
• Artificial Intelligence
• Railway Engineering

Background:

• Railway safety relies on rapid detection of internal rail defects.
• Onboard detection systems face computational resource limitations.
• Existing methods struggle to balance accuracy and speed for real-time applications.

Purpose of the Study:

• To develop an efficient and accurate AI model for onboard rail defect detection.
• To overcome the computational constraints of current railway safety systems.
• To enhance the real-time detection capabilities for high-speed rail inspection vehicles.

Main Methods:

• Designed YOLOv8n-LiteCBAM with a lightweight DepthStackNet backbone.
• Implemented model pruning and a novel Bidirectional Convolutional Block Attention Module (BiCBAM).
• Utilized ONNX Runtime for inference acceleration.

Main Results:

• Achieved 92.9% mean Average Precision (mAP) on a rail defect dataset.
• Reached inference speeds of 136.79 FPS (GPU) and 38.36 FPS (CPU).
• Outperformed other lightweight models in inference speed, meeting real-time requirements for 80 km/h inspection.

Conclusions:

• YOLOv8n-LiteCBAM offers a viable solution for efficient and accurate onboard rail defect detection.
• The model's performance supports industrial applications in expedited rail flaw detection.
• This advancement contributes to improved railway safety through enhanced detection technologies.