High-Speed Railway Intruding Object Image Generating with Generative Adversarial Networks

Area of Science:

• Computer Vision
• Artificial Intelligence
• Railway Engineering

Background:

• Foreign object intrusion poses a significant risk to high-speed railway safety.
• Accurate detection of these objects is crucial for preventing accidents.
• A scarcity of real-world intrusion samples hinders the development of effective detection methods.

Purpose of the Study:

• To develop a novel method for generating artificial railway intruding object images.
• To improve the quality and authenticity of synthetic training data for object detection systems.
• To enhance the safety and reliability of high-speed railway operations through better detection capabilities.

Main Methods:

• An improved conditional deep convolutional generative adversarial network (C-DCGAN) with a generator and multi-scale discriminators was employed.
• A novel loss function was implemented to enhance sample quality and authenticity.
• A scale estimation algorithm based on the gauge constant was proposed for realistic object scaling within railway scenes.

Main Results:

• The proposed C-DCGAN model demonstrated superior performance compared to state-of-the-art methods.
• Generated samples achieved high quality, with pixel-wise accuracy of 80.46%, mIoU of 0.65, and mAP of 0.69.
• High diversity was observed, indicated by a Fréchet-Inception Distance (FID) score of 26.87, and real-generated pedestrian pairs reached an mIoU of 0.85.

Conclusions:

• The developed C-DCGAN method effectively generates high-quality and diverse synthetic images of railway intruding objects.
• The generated data significantly aids in improving the accuracy of foreign object detection systems for high-speed railways.
• This approach offers a viable solution to the data scarcity problem, contributing to enhanced railway safety.