Cargando…

End-to-End Monocular Range Estimation for Forward Collision Warning

Estimating range to the closest object in front is the core component of the forward collision warning (FCW) system. Previous monocular range estimation methods mostly involve two sequential steps of object detection and range estimation. As a result, they are only effective for objects from specifi...

Descripción completa

Detalles Bibliográficos
Autores principales: Tang, Jie, Li, Jian
Formato: Online Artículo Texto
Lenguaje:English
Publicado: MDPI 2020
Materias:
Acceso en línea:https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7589406/
https://www.ncbi.nlm.nih.gov/pubmed/33096656
http://dx.doi.org/10.3390/s20205941
_version_ 1783600571707555840
author Tang, Jie
Li, Jian
author_facet Tang, Jie
Li, Jian
author_sort Tang, Jie
collection PubMed
description Estimating range to the closest object in front is the core component of the forward collision warning (FCW) system. Previous monocular range estimation methods mostly involve two sequential steps of object detection and range estimation. As a result, they are only effective for objects from specific categories relying on expensive object-level annotation for training, but not for unseen categories. In this paper, we present an end-to-end deep learning architecture to solve the above problems. Specifically, we represent the target range as a weighted sum of a set of potential distances. These potential distances are generated by inverse perspective projection based on intrinsic and extrinsic camera parameters, while a deep neural network predicts the corresponding weights of these distances. The whole architecture is optimized towards the range estimation task directly in an end-to-end manner with only the target range as supervision. As object category is not restricted in the training stage, the proposed method can generalize to objects with unseen categories. Furthermore, camera parameters are explicitly considered in the proposed method, making it able to generalize to images taken with different cameras and novel views. Additionally, the proposed method is not a pure black box, but provides partial interpretability by visualizing the produced weights to see which part of the image dominates the final result. We conduct experiments to verify the above properties of the proposed method on synthetic and real-world collected data.
format Online
Article
Text
id pubmed-7589406
institution National Center for Biotechnology Information
language English
publishDate 2020
publisher MDPI
record_format MEDLINE/PubMed
spelling pubmed-75894062020-10-29 End-to-End Monocular Range Estimation for Forward Collision Warning Tang, Jie Li, Jian Sensors (Basel) Article Estimating range to the closest object in front is the core component of the forward collision warning (FCW) system. Previous monocular range estimation methods mostly involve two sequential steps of object detection and range estimation. As a result, they are only effective for objects from specific categories relying on expensive object-level annotation for training, but not for unseen categories. In this paper, we present an end-to-end deep learning architecture to solve the above problems. Specifically, we represent the target range as a weighted sum of a set of potential distances. These potential distances are generated by inverse perspective projection based on intrinsic and extrinsic camera parameters, while a deep neural network predicts the corresponding weights of these distances. The whole architecture is optimized towards the range estimation task directly in an end-to-end manner with only the target range as supervision. As object category is not restricted in the training stage, the proposed method can generalize to objects with unseen categories. Furthermore, camera parameters are explicitly considered in the proposed method, making it able to generalize to images taken with different cameras and novel views. Additionally, the proposed method is not a pure black box, but provides partial interpretability by visualizing the produced weights to see which part of the image dominates the final result. We conduct experiments to verify the above properties of the proposed method on synthetic and real-world collected data. MDPI 2020-10-21 /pmc/articles/PMC7589406/ /pubmed/33096656 http://dx.doi.org/10.3390/s20205941 Text en © 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
spellingShingle Article
Tang, Jie
Li, Jian
End-to-End Monocular Range Estimation for Forward Collision Warning
title End-to-End Monocular Range Estimation for Forward Collision Warning
title_full End-to-End Monocular Range Estimation for Forward Collision Warning
title_fullStr End-to-End Monocular Range Estimation for Forward Collision Warning
title_full_unstemmed End-to-End Monocular Range Estimation for Forward Collision Warning
title_short End-to-End Monocular Range Estimation for Forward Collision Warning
title_sort end-to-end monocular range estimation for forward collision warning
topic Article
url https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7589406/
https://www.ncbi.nlm.nih.gov/pubmed/33096656
http://dx.doi.org/10.3390/s20205941
work_keys_str_mv AT tangjie endtoendmonocularrangeestimationforforwardcollisionwarning
AT lijian endtoendmonocularrangeestimationforforwardcollisionwarning