Apple orchard production estimation using deep learning strategies: A comparison of tracking-by-detection algorithms

Juan Villacrés; Michelle Viscaino; José Delpiano; Stavros Vougioukas; Fernando Auat Cheein

doi:10.1016/j.compag.2022.107513

Apple orchard production estimation using deep learning strategies: A comparison of tracking-by-detection algorithms

Juan Villacrés, Michelle Viscaino, José Delpiano, Stavros Vougioukas, Fernando Auat Cheein^*

^*Corresponding author for this work

Research output: Contribution to journal › Article › peer-review

38 Scopus citations

Abstract

The automated detection and counting of fruit in tree canopies is a key component of yield estimation systems, which are indispensable for the precision management of modern orchards. Detection and counting tasks in agricultural environments are not trivial because of challenges such as characteristics of the tree canopies, occlusion caused by leaves and the lighting conditions, among other factors. With the aim of identifying which algorithm is more suitable for yield estimation, we present a comprehensive comparison of tracking-by-detection algorithms, applied to apple counting. The tracking strategies evaluated were Kalman Filter, Kernelized Correlation Filter, Simple Online Real-Time Tracking, Multi Hypothesis Tracking, and Deep Simple Online Real-Time Tracking. The five tracking algorithms were further assessed on two novel databases constructed for this research in Multiple Object Tracking MOT format. After a sensitivity analysis of the trackers, the results show that the most robust approach is the Multiple Hypothesis Tracking, followed by the Deep Simple Online Realtime (DeepSORT), with a MOT accuracy of 97.00% and 93.00%, respectively, when having perfect detection. However, in an application case including a deep learning-based detection stage, the DeepSORT tracker obtains the lowest counting error, which on average for all videos is 20.07% and 31.52% when using YoloV5 and Faster R-CNN as detection strategies. Statistically similar results were obtained using the Kalman Filter with a counting error of 20.5% and 31.9% when detecting fruit with YoloV5 and Faster R-CNN.

Original language	English
Article number	107513
Journal	Computers and Electronics in Agriculture
Volume	204
DOIs	https://doi.org/10.1016/j.compag.2022.107513
State	Published - Jan 2023

Bibliographical note

Funding Information:
The authors would like to thank to the Advanced Center for Electrical and Electronic Engineering (AC3E) , ANID Basal project FB0008 and FONDECYT grant 1201319 . Authors would also like to thank to Universidad Técnica Federico Santa María, and ANID PFCHA/Doctorado Nacional/2020-21200684.

Funding Information:
The authors would like to thank to the Advanced Center for Electrical and Electronic Engineering (AC3E), ANID Basal project FB0008 and FONDECYT grant 1201319. Authors would also like to thank to Universidad Técnica Federico Santa María, and ANID PFCHA/Doctorado Nacional/2020-21200684.

Publisher Copyright:
© 2022 Elsevier B.V.

Keywords

Apple counting
Faster R-CNN
Multi-object tracking
Tracking-by-detection
YoloV5

Access to Document

10.1016/j.compag.2022.107513

Cite this

@article{8c0683fd58264a498e959558fa165bef,

title = "Apple orchard production estimation using deep learning strategies: A comparison of tracking-by-detection algorithms",

abstract = "The automated detection and counting of fruit in tree canopies is a key component of yield estimation systems, which are indispensable for the precision management of modern orchards. Detection and counting tasks in agricultural environments are not trivial because of challenges such as characteristics of the tree canopies, occlusion caused by leaves and the lighting conditions, among other factors. With the aim of identifying which algorithm is more suitable for yield estimation, we present a comprehensive comparison of tracking-by-detection algorithms, applied to apple counting. The tracking strategies evaluated were Kalman Filter, Kernelized Correlation Filter, Simple Online Real-Time Tracking, Multi Hypothesis Tracking, and Deep Simple Online Real-Time Tracking. The five tracking algorithms were further assessed on two novel databases constructed for this research in Multiple Object Tracking MOT format. After a sensitivity analysis of the trackers, the results show that the most robust approach is the Multiple Hypothesis Tracking, followed by the Deep Simple Online Realtime (DeepSORT), with a MOT accuracy of 97.00% and 93.00%, respectively, when having perfect detection. However, in an application case including a deep learning-based detection stage, the DeepSORT tracker obtains the lowest counting error, which on average for all videos is 20.07% and 31.52% when using YoloV5 and Faster R-CNN as detection strategies. Statistically similar results were obtained using the Kalman Filter with a counting error of 20.5% and 31.9% when detecting fruit with YoloV5 and Faster R-CNN.",

keywords = "Apple counting, Faster R-CNN, Multi-object tracking, Tracking-by-detection, YoloV5",

author = "Juan Villacr{\'e}s and Michelle Viscaino and Jos{\'e} Delpiano and Stavros Vougioukas and {Auat Cheein}, Fernando",

note = "Funding Information: The authors would like to thank to the Advanced Center for Electrical and Electronic Engineering (AC3E) , ANID Basal project FB0008 and FONDECYT grant 1201319 . Authors would also like to thank to Universidad T{\'e}cnica Federico Santa Mar{\'i}a, and ANID PFCHA/Doctorado Nacional/2020-21200684. Funding Information: The authors would like to thank to the Advanced Center for Electrical and Electronic Engineering (AC3E), ANID Basal project FB0008 and FONDECYT grant 1201319. Authors would also like to thank to Universidad T{\'e}cnica Federico Santa Mar{\'i}a, and ANID PFCHA/Doctorado Nacional/2020-21200684. Publisher Copyright: {\textcopyright} 2022 Elsevier B.V.",

year = "2023",

month = jan,

doi = "10.1016/j.compag.2022.107513",

language = "English",

volume = "204",

journal = "Computers and Electronics in Agriculture",

issn = "0168-1699",

publisher = "Elsevier B.V.",

}

TY - JOUR

T1 - Apple orchard production estimation using deep learning strategies

T2 - A comparison of tracking-by-detection algorithms

AU - Villacrés, Juan

AU - Viscaino, Michelle

AU - Delpiano, José

AU - Vougioukas, Stavros

AU - Auat Cheein, Fernando

N1 - Funding Information: The authors would like to thank to the Advanced Center for Electrical and Electronic Engineering (AC3E) , ANID Basal project FB0008 and FONDECYT grant 1201319 . Authors would also like to thank to Universidad Técnica Federico Santa María, and ANID PFCHA/Doctorado Nacional/2020-21200684. Funding Information: The authors would like to thank to the Advanced Center for Electrical and Electronic Engineering (AC3E), ANID Basal project FB0008 and FONDECYT grant 1201319. Authors would also like to thank to Universidad Técnica Federico Santa María, and ANID PFCHA/Doctorado Nacional/2020-21200684. Publisher Copyright: © 2022 Elsevier B.V.

PY - 2023/1

Y1 - 2023/1

N2 - The automated detection and counting of fruit in tree canopies is a key component of yield estimation systems, which are indispensable for the precision management of modern orchards. Detection and counting tasks in agricultural environments are not trivial because of challenges such as characteristics of the tree canopies, occlusion caused by leaves and the lighting conditions, among other factors. With the aim of identifying which algorithm is more suitable for yield estimation, we present a comprehensive comparison of tracking-by-detection algorithms, applied to apple counting. The tracking strategies evaluated were Kalman Filter, Kernelized Correlation Filter, Simple Online Real-Time Tracking, Multi Hypothesis Tracking, and Deep Simple Online Real-Time Tracking. The five tracking algorithms were further assessed on two novel databases constructed for this research in Multiple Object Tracking MOT format. After a sensitivity analysis of the trackers, the results show that the most robust approach is the Multiple Hypothesis Tracking, followed by the Deep Simple Online Realtime (DeepSORT), with a MOT accuracy of 97.00% and 93.00%, respectively, when having perfect detection. However, in an application case including a deep learning-based detection stage, the DeepSORT tracker obtains the lowest counting error, which on average for all videos is 20.07% and 31.52% when using YoloV5 and Faster R-CNN as detection strategies. Statistically similar results were obtained using the Kalman Filter with a counting error of 20.5% and 31.9% when detecting fruit with YoloV5 and Faster R-CNN.

AB - The automated detection and counting of fruit in tree canopies is a key component of yield estimation systems, which are indispensable for the precision management of modern orchards. Detection and counting tasks in agricultural environments are not trivial because of challenges such as characteristics of the tree canopies, occlusion caused by leaves and the lighting conditions, among other factors. With the aim of identifying which algorithm is more suitable for yield estimation, we present a comprehensive comparison of tracking-by-detection algorithms, applied to apple counting. The tracking strategies evaluated were Kalman Filter, Kernelized Correlation Filter, Simple Online Real-Time Tracking, Multi Hypothesis Tracking, and Deep Simple Online Real-Time Tracking. The five tracking algorithms were further assessed on two novel databases constructed for this research in Multiple Object Tracking MOT format. After a sensitivity analysis of the trackers, the results show that the most robust approach is the Multiple Hypothesis Tracking, followed by the Deep Simple Online Realtime (DeepSORT), with a MOT accuracy of 97.00% and 93.00%, respectively, when having perfect detection. However, in an application case including a deep learning-based detection stage, the DeepSORT tracker obtains the lowest counting error, which on average for all videos is 20.07% and 31.52% when using YoloV5 and Faster R-CNN as detection strategies. Statistically similar results were obtained using the Kalman Filter with a counting error of 20.5% and 31.9% when detecting fruit with YoloV5 and Faster R-CNN.

KW - Apple counting

KW - Faster R-CNN

KW - Multi-object tracking

KW - Tracking-by-detection

KW - YoloV5

UR - http://www.scopus.com/inward/record.url?scp=85144398919&partnerID=8YFLogxK

U2 - 10.1016/j.compag.2022.107513

DO - 10.1016/j.compag.2022.107513

M3 - Article

AN - SCOPUS:85144398919

SN - 0168-1699

VL - 204

JO - Computers and Electronics in Agriculture

JF - Computers and Electronics in Agriculture

M1 - 107513

ER -

Apple orchard production estimation using deep learning strategies: A comparison of tracking-by-detection algorithms

Abstract

Bibliographical note

Keywords

Access to Document

Other files and links

Fingerprint

Cite this