Assessing Deep-learning Methods for Object Detection at Sea from LWIR Images

Frederik Emil Thorsson Schöller; Martin Krarup Plenge-Feidenhans'l; Jonathan Dyssel Stets; Mogens Blanke

doi:10.1016/j.ifacol.2019.12.284

Assessing Deep-learning Methods for Object Detection at Sea from LWIR Images

Frederik Emil Thorsson Schöller, Martin Krarup Plenge-Feidenhans'l, Jonathan Dyssel Stets, Mogens Blanke

Research output: Contribution to journal › Journal article › Research › peer-review

432 Downloads (Pure)

Abstract

This paper assesses the performance of three convolutional neural networks for object detection at sea using Long Wavelength Infrared (LWIR) images in the 8 − 14µm range. Capturing images from ferries and annotating 20k images, fine-tuning is done of three state of art deep neural networks: RetinaNet, YOLO and Faster R-CNN. Targeting on vessels and buoys as two main classes of interest for navigation, performance is quantified by the cardinality of true and false positives and negatives in a random validation set. Calculating precision and recall as functions of tuning parameters for the three classifiers, noticeable differences are found between the three networks when used for LWIR image object classification at sea. The results lead to conclusions on imaging requirements when classification is used to support navigation.

Original language	English
Book series	I F A C Workshop Series
Volume	52
Issue number	21
Pages (from-to)	64-71
ISSN	1474-6670
DOIs	https://doi.org/10.1016/j.ifacol.2019.12.284
Publication status	Published - 2019
Event	12th Control Applications in Marine Systems, Robotics, and Vehicles - Chunghi & Byiung Jun Park KI Building, KAIST, Daejeon, Korea, Republic of Duration: 18 Sep 2019 → 20 Sep 2019 Conference number: 12 http://cams-wroco.org/

Conference

Conference	12th Control Applications in Marine Systems, Robotics, and Vehicles
Number	12
Location	Chunghi & Byiung Jun Park KI Building, KAIST
Country	Korea, Republic of
City	Daejeon
Period	18/09/2019 → 20/09/2019
Internet address	http://cams-wroco.org/

Keywords

Object Detection
Autonomous marine crafts
Navigation
Long-wave Infra-red
Detection at sea
Autonomous Ship

Access to Document

10.1016/j.ifacol.2019.12.284

Cams2019 fs mf js mbAccepted author manuscript, 1.17 MB

Fingerprint Dive into the research topics of 'Assessing Deep-learning Methods for Object Detection at Sea from LWIR Images'. Together they form a unique fingerprint.

2 Active

ShippingLab Autonomy
Blanke, M., Galeazzi, R., Dittmann, K., Hansen, S., Papageorgiou, D., Nalpantidis, L., Schöller, F. E. T., Plenge-Feidenhans'l, M. K., Hansen, P. N., Andersen, R. H., Becktor, J. B., Enevoldsen, T. T., Dagdilelis, D., Karstensen, P. I. H., Nielsen, R. E., Garde, J. & Ravn, O.
01/04/2019 → 31/12/2022
Project: Research
ASAN: Autonomous Situation Awareness for Navigation
Blanke, M., Hansen, S., Papageorgiou, D. & Stets, J. D.
01/12/2018 → 31/03/2021
Project: Research

Cite this

@article{2d797198170742e38de07eab18f2d360,

title = "Assessing Deep-learning Methods for Object Detection at Sea from LWIR Images",

abstract = "This paper assesses the performance of three convolutional neural networks for object detection at sea using Long Wavelength Infrared (LWIR) images in the 8 − 14µm range. Capturing images from ferries and annotating 20k images, fine-tuning is done of three state of art deep neural networks: RetinaNet, YOLO and Faster R-CNN. Targeting on vessels and buoys as two main classes of interest for navigation, performance is quantified by the cardinality of true and false positives and negatives in a random validation set. Calculating precision and recall as functions of tuning parameters for the three classifiers, noticeable differences are found between the three networks when used for LWIR image object classification at sea. The results lead to conclusions on imaging requirements when classification is used to support navigation.",

keywords = "Object Detection, Autonomous marine crafts, Navigation, Long-wave Infra-red, Detection at sea, Autonomous Ship",

author = "Sch{\"o}ller, {Frederik Emil Thorsson} and Plenge-Feidenhans'l, {Martin Krarup} and Stets, {Jonathan Dyssel} and Mogens Blanke",

year = "2019",

doi = "10.1016/j.ifacol.2019.12.284",

language = "English",

volume = "52",

pages = "64--71",

journal = "I F A C Workshop Series",

issn = "1474-6670",

publisher = "Elsevier",

number = "21",

note = "12th Control Applications in Marine Systems, Robotics, and Vehicles, IFAC CAMS 2019 ; Conference date: 18-09-2019 Through 20-09-2019",

url = "http://cams-wroco.org/",

}

TY - JOUR

T1 - Assessing Deep-learning Methods for Object Detection at Sea from LWIR Images

AU - Schöller, Frederik Emil Thorsson

AU - Plenge-Feidenhans'l, Martin Krarup

AU - Stets, Jonathan Dyssel

AU - Blanke, Mogens

N1 - Conference code: 12

PY - 2019

Y1 - 2019

N2 - This paper assesses the performance of three convolutional neural networks for object detection at sea using Long Wavelength Infrared (LWIR) images in the 8 − 14µm range. Capturing images from ferries and annotating 20k images, fine-tuning is done of three state of art deep neural networks: RetinaNet, YOLO and Faster R-CNN. Targeting on vessels and buoys as two main classes of interest for navigation, performance is quantified by the cardinality of true and false positives and negatives in a random validation set. Calculating precision and recall as functions of tuning parameters for the three classifiers, noticeable differences are found between the three networks when used for LWIR image object classification at sea. The results lead to conclusions on imaging requirements when classification is used to support navigation.

AB - This paper assesses the performance of three convolutional neural networks for object detection at sea using Long Wavelength Infrared (LWIR) images in the 8 − 14µm range. Capturing images from ferries and annotating 20k images, fine-tuning is done of three state of art deep neural networks: RetinaNet, YOLO and Faster R-CNN. Targeting on vessels and buoys as two main classes of interest for navigation, performance is quantified by the cardinality of true and false positives and negatives in a random validation set. Calculating precision and recall as functions of tuning parameters for the three classifiers, noticeable differences are found between the three networks when used for LWIR image object classification at sea. The results lead to conclusions on imaging requirements when classification is used to support navigation.

KW - Object Detection

KW - Autonomous marine crafts

KW - Navigation

KW - Long-wave Infra-red

KW - Detection at sea

KW - Autonomous Ship

U2 - 10.1016/j.ifacol.2019.12.284

DO - 10.1016/j.ifacol.2019.12.284

M3 - Journal article

VL - 52

SP - 64

EP - 71

JO - I F A C Workshop Series

JF - I F A C Workshop Series

SN - 1474-6670

IS - 21

T2 - 12th Control Applications in Marine Systems, Robotics, and Vehicles

Y2 - 18 September 2019 through 20 September 2019

ER -

Assessing Deep-learning Methods for Object Detection at Sea from LWIR Images

Abstract

Conference

Keywords

Access to Document

ShippingLab Autonomy

ASAN: Autonomous Situation Awareness for Navigation

Cite this