Kafetzis, Dimitrios (Athens University of Economics and Business), Fourfouris, Ioannis (Athens University of Economics and Business), Argyropoulos, Savvas (StreamOwl), Koutsopoulos, Iordanis (Athens University of Economics and Business)

UAV-Assisted Aerial Survey of Railways Using Deep Learning

Scheduled for presentation during the Regular Session "UAS Applications V" (FrB3), Friday, September 4, 2020, 13:10−13:30, Edessa

2020 International Conference on Unmanned Aircraft Systems (ICUAS), September 1-4, 2020 (Postponed from June 9-12, 2020), Athens, Greece

This information is tentative and subject to change. Compiled on April 25, 2024

Abstract

Unmanned Air Vehicles (UAVs) are currently in use for diverse applications such as critical infrastructure monitoring. Monitoring is based on video capture by a video camera and subsequent use of Deep Learning (DL) techniques to perform image recognition. In this paper we do a proof-of-concept validation of DL and UAV-assisted monitoring for vegetation and object detection on railways. The large diversity of vegetation, terrain and railway settings increases the challenges for object detection and classification. Moreover, the creation of an appropriate dataset for the training of a classifier is a nontrivial task per se. We show the related challenges in this setting, and we create from scratch a dedicated dataset with manual annotation for vegetation management on railways, based on publicly available video clips and our own video recordings at a railway. To the best of our knowledge this is the first dedicated dataset for this application. Next, we develop a DL pipeline on this dataset and evaluate its performance for different classes of vegetation and obstacles on the railways. Our approach leads to satisfactory detection accuracy, especially given the diversity of obstacles and the fact that most objects to be detected appear at the background of the frame image. Also, we test our classifier models in the NVIDIA Jetson Nano platform, which is the on-board computing system of a UAV that we used for on-site testing. The classifier may operate on the Jetson Nano board presenting a good and viable performance.