A GAN-based Deep Learning Model for the Object Detection of Autonomous High-Speed Trains

With the development of autonomous trains that are operated without human labor in high-speed railways, the realtime detection of objects with computer vision techniques, including signals, falling rocks, and the infrastructure that collapses near the railroad, has received increased attention. Although existing studies have proposed different deep learning-based models, the critical issue is the lack of real-world data samples since the accidents caused by these objects seldom happen in the practical operations of high-speed trains. In this paper, we propose a deep-learning-based Generative Adversarial Network (GAN) framework for the realtime detection of objects with cameras in front of the running train. In particular, our GAN framework employs two connected deep-learning networks. The first network adopts a pix2pix scheme to automatically extend the few-shot data, i.e., data samples with objects (signals, obstacles, etc.) and noises. Our second network, i.e., the detection network, aims to detect and classify these objects precisely. We conduct experiments and compare our GAN with a few benchmarks on real-world data sets. The results reveal that our proposed GAN model significantly outperforms the commonly used deep learning models in detecting objects.