Deep learning approach to vehicle pose estimation from polarimetric image data

Matthew D. Siefring, B. Ratliff, Brett S. Ballard
{"title":"Deep learning approach to vehicle pose estimation from polarimetric image data","authors":"Matthew D. Siefring, B. Ratliff, Brett S. Ballard","doi":"10.1117/12.2677123","DOIUrl":null,"url":null,"abstract":"Object pose estimation is an important problem in the field of remote sensing that provides valuable information for target identification tasks. Polarization is a fundamental property of light that contains useful information about the physical properties of an object, such as shape and surface material properties. Polarization imaging has been shown to have advantages over conventional imaging techniques for object detection and feature extraction in a variety of challenging scenarios, including low light, high background clutter, and low visibility conditions. In this work, we investigate using polarimetric imaging to improve the performance of deep learning approaches to object pose estimation on a range of model target vehicles. We collect polarimetric imaging data and labeled ground truth pose data on the target vehicles in a controlled solar simulation laboratory environment under precise sensor, object, and solar source geometries. We first establish baseline performance of our approach by training our network using conventional visible RGB s0 images under favorable lighting conditions. We then make use of the full linear Stokes images for each color channel in various configurations, retrain our network, and compare performance. We furthermore propose an ensemble method to combine features obtained from convolutional neural networks trained on both conventional RGB and Stokes-vector images. These obtained ensemble features are then used to train a multi-layer perceptron. Experimental results demonstrate that combining polarization imaging with conventional imaging can improve feature extraction and the accuracy of deep learning-based approaches to pose estimation.","PeriodicalId":434863,"journal":{"name":"Optical Engineering + Applications","volume":"24 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2023-10-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Optical Engineering + Applications","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1117/12.2677123","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 0

Abstract

Object pose estimation is an important problem in the field of remote sensing that provides valuable information for target identification tasks. Polarization is a fundamental property of light that contains useful information about the physical properties of an object, such as shape and surface material properties. Polarization imaging has been shown to have advantages over conventional imaging techniques for object detection and feature extraction in a variety of challenging scenarios, including low light, high background clutter, and low visibility conditions. In this work, we investigate using polarimetric imaging to improve the performance of deep learning approaches to object pose estimation on a range of model target vehicles. We collect polarimetric imaging data and labeled ground truth pose data on the target vehicles in a controlled solar simulation laboratory environment under precise sensor, object, and solar source geometries. We first establish baseline performance of our approach by training our network using conventional visible RGB s0 images under favorable lighting conditions. We then make use of the full linear Stokes images for each color channel in various configurations, retrain our network, and compare performance. We furthermore propose an ensemble method to combine features obtained from convolutional neural networks trained on both conventional RGB and Stokes-vector images. These obtained ensemble features are then used to train a multi-layer perceptron. Experimental results demonstrate that combining polarization imaging with conventional imaging can improve feature extraction and the accuracy of deep learning-based approaches to pose estimation.
基于偏振图像数据的车辆姿态估计的深度学习方法
目标位姿估计是遥感领域的一个重要问题,为目标识别任务提供了有价值的信息。偏振是光的基本特性,它包含了物体物理特性的有用信息,如形状和表面材料特性。在各种具有挑战性的情况下,包括低光、高背景杂波和低能见度条件下,极化成像在目标检测和特征提取方面比传统成像技术具有优势。在这项工作中,我们研究了使用偏振成像来提高深度学习方法在一系列模型目标车辆上的目标姿态估计的性能。我们在受控的太阳模拟实验室环境中,在精确的传感器、物体和太阳光源几何形状下,收集目标车辆的偏振成像数据和标记的地面真态数据。我们首先通过在有利的照明条件下使用常规可见RGB 50图像训练我们的网络来建立我们方法的基线性能。然后,我们在各种配置下对每个颜色通道使用全线性Stokes图像,重新训练我们的网络,并比较性能。我们进一步提出了一种集成方法,将在传统RGB和stokes矢量图像上训练的卷积神经网络获得的特征结合起来。然后使用这些获得的集成特征来训练多层感知器。实验结果表明,将偏振成像与常规成像相结合可以提高基于深度学习的姿态估计方法的特征提取和精度。
本文章由计算机程序翻译,如有差异,请以英文原文为准。
求助全文
约1分钟内获得全文 求助全文
来源期刊
自引率
0.00%
发文量
0
×
引用
GB/T 7714-2015
复制
MLA
复制
APA
复制
导出至
BibTeX EndNote RefMan NoteFirst NoteExpress
×
提示
您的信息不完整,为了账户安全,请先补充。
现在去补充
×
提示
您因"违规操作"
具体请查看互助需知
我知道了
×
提示
确定
请完成安全验证×
copy
已复制链接
快去分享给好友吧!
我知道了
右上角分享
点击右上角分享
0
联系我们:info@booksci.cn Book学术提供免费学术资源搜索服务,方便国内外学者检索中英文文献。致力于提供最便捷和优质的服务体验。 Copyright © 2023 布克学术 All rights reserved.
京ICP备2023020795号-1
ghs 京公网安备 11010802042870号
Book学术文献互助
Book学术文献互助群
群 号:604180095
Book学术官方微信