Application of Computer Vision and Deep Learning in the railway domain for autonomous train stop operation

Abstract

The purpose of this paper is to present the results of the analysis of the application of Deep Learning in the railway domain with a particular focus on a train stop operation. The paper proposes an approach consisting of monocular vision-based and Deep Learning architectures. Even the difficulties imposed by actual regulation, the findings show that Deep Learning architecture can offer promising results in railway localization using techniques like visual odometry, SLAM or pose estimation. Besides, in spite of the many datasets available in the literature needed to train the neural network, none of them have been created for indoor railway environments. Therefore, a new dataset should be created. Furthermore, the paper presents future research and development suggestions for railway applications which contribute to guiding the mid-term research and development.