Railway Object Detection

Project overview

Railway infrastructure images were annotated for a computer vision dataset focused on detecting and identifying the different objects present in railway environments. The project involved locating railway-related objects and creating accurate annotations for each class to generate high-quality training data for an AI-based object detection model.

The dataset included a variety of objects such as coins, bolts, traffic signs, hectometer stones and cable poles commonly found around railway environments.

Work performed

• Reviewed and organised the railway image dataset before annotation.
• All target objects present in railway environments were identified.
• Accurate bounding boxes were created around individual objects.
• Consistent labelling was maintained across all five classes.
• Target objects were carefully distinguished from tracks, equipment, vegetation and surrounding backgrounds.
• Annotated partially visible objects where applicable.
• Objects appearing at different sizes, distances and viewing angles were handled individually.
• Quality checks identified missing, incorrect or poorly positioned annotations.
• Errors were corrected and consistency maintained throughout the dataset.
• Confirmed the final annotations were suitable for training an object detection model.

What made it hard

• Detecting small objects such as coins and bolts in railway environments.
• Railway objects had to be distinguished from similar-looking background elements.
• Partially visible or occluded objects complicated box placement.
• Traffic signs appeared at different distances and orientations.
• Hectometer stones required careful identification among surrounding infrastructure.
• Cable poles and other tall objects needed accurate bounding box placement.
• Complex backgrounds containing tracks, gravel, vegetation and equipment made detection challenging.
• Keeping annotations accurate across multiple object categories required careful quality control.

Result

A high-quality railway object detection dataset was completed with accurate and consistent bounding box annotations for coins, bolts, traffic signs, hectometer stones and cable poles. The dataset was prepared for training a computer vision model able to detect and identify railway-related objects automatically in real-world railway environments.

Final deliverable

Accurate and consistently labelled bounding box annotations were delivered for all five railway object classes. The final dataset met the project requirements and was ready for AI model training and automated railway infrastructure monitoring and object detection applications.