Flooded Road Detection from Driving Recorder: Training Deep Net for Rare Event Using GANs Semantic Information

Sho Nakamura, Shintaro Ono, Hiroshi Kawasaki

Research output: Contribution to journalArticlepeer-review

Abstract

It is important for traffic management to understand unusual conditions or road abnormalities caused by natural disasters (such as an earthquake or heavy rain) or traffic congestion caused by special events (such as festivals at tourist spots). Among these, we focused on flooded roads as unusual events and proposed a method to detect it automatically, using deep-learning methods from driving videos. Because such unusual events rarely occur, the amount of training data for deep learning is usually insufficient. Therefore, we propose a data-augmentation approach using Generative Adversarial Networks (GANs) to solve the problem. To effectively augment the data, we propose a multi-domain image-to-image transformation by GANs. In addition, to increase the robustness of the image transformation, we newly introduce semantic information. We synthesized a new dataset using GANs and verified the performance of our method by detecting flooded scenes.

Original languageEnglish
JournalInternational Journal of Intelligent Transportation Systems Research
Volume19
Issue number1
DOIs
Publication statusPublished - Apr 2021

All Science Journal Classification (ASJC) codes

  • Software
  • Control and Systems Engineering
  • Neuroscience(all)
  • Information Systems
  • Automotive Engineering
  • Aerospace Engineering
  • Computer Science Applications
  • Applied Mathematics

Fingerprint Dive into the research topics of 'Flooded Road Detection from Driving Recorder: Training Deep Net for Rare Event Using GANs Semantic Information'. Together they form a unique fingerprint.

Cite this