Embedded object detection applying Deep Neural Networks in railway domain

Abstract

In the last few years, research on deep learning application on the transportation industry has grown. One of the tasks afforded on those works is the object detection, a essential function in autonomous vehicles, including railway vehicles. While the application of deep learning for object detection is increasing in railway domain, proposed methods have to be yet tested on embedded hardware. This work explores the efficiency of the standard YoloV3 detector embedded on a NVIDIA Jetson AGX Xavier to infer traffic signals in the railway domain. Furthermore, different architectures of YoloV3 are analyzed and compared to find the best output for the used dataset. A data augmentation technique called RICAP-DET is developed to create the training dataset by generating labeled images from cutouts of a set of images. The results show that YoloV3 can be used to detect rail traffic-signals in real time on an embedded platform and that RICAP-DET is adequate to train YoloV3.