Detail publikace

Convolutional Neural Networks for the Odometry Estimation

VEĽAS, M. ŠPANĚL, M. HRADIŠ, M. HEROUT, A.

Originální název

Convolutional Neural Networks for the Odometry Estimation

Typ

článek v časopise - ostatní, Jost

Jazyk

angličtina

Originální abstrakt

This article presents a novel method for odometry estimation from 3D data of Velodyne LiDAR scanner using convolutional neural networks. For training and forward evaluation of the proposed networks, the original data is encoded into 2D matrices. In experiments with the KITTI dataset, our networks show significantly higher accuracy in estimation of the translational motion parameters compared to the state of the art LOAM method. In addition, they achieve higher speed and real-time performance. Using data provided by the IMU sensor, it is possible to estimate odometry and align the point cloud with a high precision. The proposed method can replace the odometry estimation from the wheel encoders or supplement the missing GPS data when the GNSS signal is not available (for example, during the interior mapping). In addition, we propose alternate CNNs for the estimation of the rotational motion that achieve results comparable to the state of the art. Our solution delivers real-time performance and accuracy to provide online preview of the mapping and to verify the completeness of the map during the mission.

Klíčová slova

Odometry, Velodyne, LiDAR, CNN, KITTI

Autoři

VEĽAS, M.; ŠPANĚL, M.; HRADIŠ, M.; HEROUT, A.

Vydáno

24. 1. 2019

ISSN

0921-0296

Periodikum

Journal of Intelligent and Robotics Systems

Ročník

2019

Číslo

93

Stát

Nizozemsko

Strany od

1

Strany do

22

Strany počet

22

URL

BibTex

@article{BUT162266,
  author="Martin {Veľas} and Michal {Španěl} and Michal {Hradiš} and Adam {Herout}",
  title="Convolutional Neural Networks for the Odometry Estimation",
  journal="Journal of Intelligent and Robotics Systems",
  year="2019",
  volume="2019",
  number="93",
  pages="1--22",
  issn="0921-0296",
  url="https://www.fit.vut.cz/research/publication/11763/"
}