An Online Rail Track Fastener Classification System Based on YOLO Models.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Railway Engineering

Background:

• Manual rail track inspection is labor-intensive and prone to human error.
• The need for automated, efficient, and accurate defect detection systems is critical for railway safety.

Purpose of the Study:

• To develop a real-time computer vision system for identifying rail track defects.
• To leverage the YOLOv4-Tiny neural network for automated defect recognition.

Main Methods:

• Utilized YOLOv4-Tiny for real-time defect identification, covering fasteners, rail surfaces, sleepers, and rail waist.
• Implemented a system with a high-performance notebook, sports cameras, and parallel processing on a moving cart (30 km/h).
• Employed Cycle Generative Adversarial Network (GAN) for data augmentation, increasing the dataset to 3800 (upward) and 967 (sideward) images.

Main Results:

• YOLOv4-Tiny achieved high performance: 91.7% mAP, 92% precision, and 91% recall for upward defects.
• For sideward defects, YOLOv4-Tiny demonstrated superior results with 99.16% mAP, 96% precision, and 94% recall.
• The system processed images at 150 FPS, enabling efficient real-time inspection.

Conclusions:

• The YOLOv4-Tiny based system offers an effective solution for automated rail track inspection.
• The developed methodology significantly enhances the efficiency and accuracy of defect detection, reducing manpower requirements.