ROBUST MONOCULAR LOCALIZATION OF DRONES BY ADAPTING DOMAIN MAPS TO DEPTH PREDICTION INACCURACIES Priyesh Shukla Sureshkumar S. Alex C. Stutts Sathya Ravi Theja Tulabandhula and Amit R. Trivedi

2025-05-03 0 0 1.53MB 5 页 10玖币

侵权投诉

ROBUST MONOCULAR LOCALIZATION OF DRONES BY ADAPTING DOMAIN MAPS TO

DEPTH PREDICTION INACCURACIES

Priyesh Shukla, Sureshkumar S., Alex C. Stutts, Sathya Ravi, Theja Tulabandhula, and Amit R. Trivedi

University of Illinois at Chicago, USA

ABSTRACT

We present a novel monocular localization framework by

jointly training deep learning-based depth prediction and

Bayesian ﬁltering-based pose reasoning. The proposed cross-

modal framework signiﬁcantly outperforms deep learning-

only predictions with respect to model scalability and tol-

erance to environmental variations. Speciﬁcally, we show

little-to-no degradation of pose accuracy even with extremely

poor depth estimates from a lightweight depth predictor. Our

framework also maintains high pose accuracy in extreme

lighting variations compared to standard deep learning, even

without explicit domain adaptation. By openly representing

the map and intermediate feature maps (such as depth es-

timates), our framework also allows for faster updates and

reusing intermediate predictions for other tasks, such as ob-

stacle avoidance, resulting in much higher resource efﬁciency.

Index Terms—Depth neural network, drone localization.

1 Introduction

For self-navigation, the most fundamental computation re-

quired for a vehicle is to determine its position and orienta-

tion, i.e., pose during motion. Higher-level path planning ob-

jectives such as motion tracking and obstacle avoidance oper-

ate by continuously estimating vehicle’s pose. Recently, deep

neural networks (DNNs) have shown a remarkable ability for

vision-based pose estimation in highly complex and cluttered

environments [1–3]. For visual pose estimation, DNNs can

learn the correlation of vehicle’s position/orientation and vi-

sual ﬁelds to a mounted camera. Thereby, vehicle’s pose can

be predicted using a monocular camera alone. In contrast, the

traditional methods required bulky and power-hungry range

sensors or stereo vision sensors to resolve the ambiguity be-

tween an object’s distance and its scale [4, 5].

However, DNN’s implicit learning of ﬂying domain fea-

tures such as its map, placement of objects, coordinate frame,

domain structure, etc. in a standard pose-DNN also affects

the robustness and adaptability of pose estimations. The tra-

ditional ﬁltering-based approaches [6] account for the ﬂying

space structure using explicit representations such as voxel

grids, occupancy grid, Gaussian mixture model (GMM), etc.

[7]; thereby, updates to the ﬂying space such as map exten-

sion, new objects, and locations can be more easily accommo-

dated. Comparatively, DNN-based estimators cannot handle

selective map updates, and the entire model must be retrained

even under small randomized or structured perturbations. Ad-

ditionally, ﬁltering loops in traditional methods can adjudi-

cate predictive uncertainties against measurements to system-

atically prune hypothesis space and can express prediction

conﬁdence along with the prediction itself [8]. Whereas feed-

forward pose estimations from a deterministic DNN are vul-

nerable to measurement and modeling uncertainties.

In thie paper, we use integrate traditional ﬁltering tech-

niques with deep learning to overcome such limitations of

DNN-based pose estimation while exploiting their suitability

to operate efﬁciently with monocular cameras alone. Specif-

ically, we present a novel framework for visual localization

by integrating DNN-based depth prediction and Bayesian

ﬁltering-based pose localization. In Figure 1, avoiding range

sensors for localization, we utilize a DNN-based lightweight

depth prediction network at the front end and sequential

Bayesian estimation at the back end. Our key observation

is that, unlike pose estimation, which innately depends on

map characteristics such as spatial structure, objects, coordi-

nate frame, etc., depth prediction is map-independent [9, 10].

Thus, by applying deep learning only on domain-independent

tasks and utilizing traditional models where domain is openly

(or explicitly) represented helps improve the predictive ro-

bustness. Limiting deep learning to only domain-independent

tasks also allows our framework to utilize vast training sets

from unrelated domains. Open representation of map and

depth estimates enables faster domain-speciﬁc updates and

utilization of intermediate feature maps for other autonomy

objectives, such as obstacle avoidance, thus improving com-

putational efﬁciency.

2 Monocular Localization with Depth

Neural Network and Pose Filters

In Figure 1, our framework integrates deep learning-based

depth prediction and Bayesian ﬁlters for visual pose local-

ization in the 3D space. At the front end, a depth DNN scans

monocular camera images to predict the relative depth of im-

age pixels from the camera’s focal point. A particle ﬁlter lo-

arXiv:2210.15559v1 [cs.CV] 27 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ROBUSTMONOCULARLOCALIZATIONOFDRONESBYADAPTINGDOMAINMAPSTODEPTHPREDICTIONINACCURACIESPriyeshShukla,SureshkumarS.,AlexC.Stutts,SathyaRavi,ThejaTulabandhula,andAmitR.TrivediUniversityofIllinoisatChicago,USAABSTRACTWepresentanovelmonocularlocalizationframeworkbyjointlytrainingdeeplearning-baseddepthpred...

展开>> 收起<<

ROBUST MONOCULAR LOCALIZATION OF DRONES BY ADAPTING DOMAIN MAPS TO DEPTH PREDICTION INACCURACIES Priyesh Shukla Sureshkumar S. Alex C. Stutts Sathya Ravi Theja Tulabandhula and Amit R. Trivedi.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

ROBUST MONOCULAR LOCALIZATION OF DRONES BY ADAPTING DOMAIN MAPS TO DEPTH PREDICTION INACCURACIES Priyesh Shukla Sureshkumar S. Alex C. Stutts Sathya Ravi Theja Tulabandhula and Amit R. Trivedi

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: