AI System for Train Display Data Extraction

Computer Vision • OCR • YOLOv11 Detection • Automation Pipeline

Project Overview

The client recorded 4.000–12.000 images per train ride, capturing the onboard energy and traction display every 1–3 seconds. Manual extraction was impossible due to blurry frames, changing lighting (tunnels, reflections, shadows), autofocus issues and tilted perspective.

The goal was to build an automated AI system capable of reading six key indicators from these screen images

Business Problem

Traditional OCR failed because many values were not text (bar levels, triangle indicators, gauge needles). A vision model was required to detect bars, segments and geometric markers not just read text.

The dataset suffered from:

• Dark tunnel frames
• Strong reflections on the glass
• Shadow gradients
• Motion blur
• Camera shake and slight rotation
• Autofocus instability
• Inconsistent brightness and contrast

Solution

I developed a two-stage extraction pipeline combining YOLOv11-M, OCR and AWS cloud processing:

YOLOv11-M Model for Visual Element Detection
OCR + AWS for Numeric Values
Data Structuring into Excel

Project Milestones

YOLOv11-M Training for Visual Data Extraction

Complete annotation and model training of YOLOv11-M to detect all visual elements on the train display: bar diagrams, bar peaks, traction direction (blue/red), current indicator triangle and gauge areas.

The model was optimized to handle blur, reflections, dark tunnels, rotation and unstable focus.

AWS Textract Pipeline & Excel Data Output System

Development of a robust OCR pipeline using AWS Textract to extract numeric fields such as energy consumption, recuperation, current and speed.

Includes text cleaning, normalization, domain-specific corrections and an automated application that exports all processed data directly into a structured Excel file ready for analysis.

Project Conclusion

This project delivers a fully automated system for extracting and interpreting data from train-display images using advanced OCR, YOLOv11-M visual detection and structured Excel generation.
The combined pipeline achieves near-100% accuracy, even under challenging conditions such as blur, reflections, dark tunnels, tilted frames and inconsistent lighting.

The system was tested extensively on new datasets provided by the client, confirming stable performance and strong generalization on unseen images.
Because of the high accuracy and reliability, the client immediately requested additional similar projects, confirming the success and real-world value of the solution.

Impact

• Reliable extraction of all train-display metrics, even from low-quality or unstable images

• Near-perfect accuracy thanks to a custom-trained YOLOv11-M model and OCR validation steps

• Major time savings for the client, eliminating manual review of thousands of images

• Scalable pipeline ready for future expansions, new indicators and automation features

• Project quality led the client to commission multiple follow-up orders

⭐ Client Reviews

Page updated

Google Sites

Report abuse