A Dual Scale Matching Model for Long-Term Association

Zhen Ye; Yubin Wu; Shuai Wang; Yang Zhang; Yanbing Chen; Wei Ke; Hao Sheng

doi:10.1007/978-3-030-59016-1_54

A Dual Scale Matching Model for Long-Term Association

Zhen Ye, Yubin Wu, Shuai Wang, Yang Zhang, Yanbing Chen, Wei Ke, Hao Sheng

Faculty of Applied Sciences

Beihang University

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

1 Citation (Scopus)

Abstract

Multi-object tracking can be characterized as a data association problem. The advantage of RNN in processing temporal dependence makes it an ideal selection in data association. When factors such as scene congestion and weak illumination cause detection failure especially long intervals, association is often very difficult and eventually leads to tracking failure. To solve this problem, Dual Scale Matching model (DSM) containing a Motion Trend Match Network (MTMNet) and an Appearance History Memory Network (AHMNet) is proposed. DSM is a long-term optimization method based on multiple hypothesis tracking. MTMNet aims to learn a similarity metric matching function between tracklets leveraging the motion feature. AHMNet is designed to provide optimal pruning strategies leveraging long period appearance feature. Our method is effective on MOT17 benchmark and it shows that we achieve considerable competitive results with current state-of-the-art trackers.

Original language	English
Title of host publication	Wireless Algorithms, Systems, and Applications - 15th International Conference, WASA 2020, Proceedings
Editors	Dongxiao Yu, Falko Dressler, Jiguo Yu
Publisher	Springer Science and Business Media Deutschland GmbH
Pages	653-665
Number of pages	13
ISBN (Print)	9783030590154
DOIs	https://doi.org/10.1007/978-3-030-59016-1_54
Publication status	Published - 2020
Event	15th International Conference on Wireless Algorithms, Systems, and Applications, WASA 2020 - Qingdao, China Duration: 13 Sept 2020 → 15 Sept 2020

Publication series

Name	Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume	12384 LNCS
ISSN (Print)	0302-9743
ISSN (Electronic)	1611-3349

Conference

Conference	15th International Conference on Wireless Algorithms, Systems, and Applications, WASA 2020
Country/Territory	China
City	Qingdao
Period	13/09/20 → 15/09/20

Keywords

Appearance History Memory Net
Data association
Dual Scale Matching
Motion Trend Match Net
Multi-object tracking

Access to Document

10.1007/978-3-030-59016-1_54

Cite this

Ye, Z., Wu, Y., Wang, S., Zhang, Y., Chen, Y., Ke, W., & Sheng, H. (2020). A Dual Scale Matching Model for Long-Term Association. In D. Yu, F. Dressler, & J. Yu (Eds.), Wireless Algorithms, Systems, and Applications - 15th International Conference, WASA 2020, Proceedings (pp. 653-665). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 12384 LNCS). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-030-59016-1_54

Ye, Zhen ; Wu, Yubin ; Wang, Shuai et al. / A Dual Scale Matching Model for Long-Term Association. Wireless Algorithms, Systems, and Applications - 15th International Conference, WASA 2020, Proceedings. editor / Dongxiao Yu ; Falko Dressler ; Jiguo Yu. Springer Science and Business Media Deutschland GmbH, 2020. pp. 653-665 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).

@inproceedings{544e9d73c1744f0e8beeab5fabd40b6f,

title = "A Dual Scale Matching Model for Long-Term Association",

abstract = "Multi-object tracking can be characterized as a data association problem. The advantage of RNN in processing temporal dependence makes it an ideal selection in data association. When factors such as scene congestion and weak illumination cause detection failure especially long intervals, association is often very difficult and eventually leads to tracking failure. To solve this problem, Dual Scale Matching model (DSM) containing a Motion Trend Match Network (MTMNet) and an Appearance History Memory Network (AHMNet) is proposed. DSM is a long-term optimization method based on multiple hypothesis tracking. MTMNet aims to learn a similarity metric matching function between tracklets leveraging the motion feature. AHMNet is designed to provide optimal pruning strategies leveraging long period appearance feature. Our method is effective on MOT17 benchmark and it shows that we achieve considerable competitive results with current state-of-the-art trackers.",

keywords = "Appearance History Memory Net, Data association, Dual Scale Matching, Motion Trend Match Net, Multi-object tracking",

author = "Zhen Ye and Yubin Wu and Shuai Wang and Yang Zhang and Yanbing Chen and Wei Ke and Hao Sheng",

note = "Publisher Copyright: {\textcopyright} 2020, Springer Nature Switzerland AG.; 15th International Conference on Wireless Algorithms, Systems, and Applications, WASA 2020 ; Conference date: 13-09-2020 Through 15-09-2020",

year = "2020",

doi = "10.1007/978-3-030-59016-1_54",

language = "English",

isbn = "9783030590154",

series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",

publisher = "Springer Science and Business Media Deutschland GmbH",

pages = "653--665",

editor = "Dongxiao Yu and Falko Dressler and Jiguo Yu",

booktitle = "Wireless Algorithms, Systems, and Applications - 15th International Conference, WASA 2020, Proceedings",

address = "Germany",

}

Ye, Z, Wu, Y, Wang, S, Zhang, Y, Chen, Y, Ke, W & Sheng, H 2020, A Dual Scale Matching Model for Long-Term Association. in D Yu, F Dressler & J Yu (eds), Wireless Algorithms, Systems, and Applications - 15th International Conference, WASA 2020, Proceedings. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 12384 LNCS, Springer Science and Business Media Deutschland GmbH, pp. 653-665, 15th International Conference on Wireless Algorithms, Systems, and Applications, WASA 2020, Qingdao, China, 13/09/20. https://doi.org/10.1007/978-3-030-59016-1_54

A Dual Scale Matching Model for Long-Term Association. / Ye, Zhen; Wu, Yubin; Wang, Shuai et al.
Wireless Algorithms, Systems, and Applications - 15th International Conference, WASA 2020, Proceedings. ed. / Dongxiao Yu; Falko Dressler; Jiguo Yu. Springer Science and Business Media Deutschland GmbH, 2020. p. 653-665 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 12384 LNCS).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - A Dual Scale Matching Model for Long-Term Association

AU - Ye, Zhen

AU - Wu, Yubin

AU - Wang, Shuai

AU - Zhang, Yang

AU - Chen, Yanbing

AU - Ke, Wei

AU - Sheng, Hao

PY - 2020

Y1 - 2020

N2 - Multi-object tracking can be characterized as a data association problem. The advantage of RNN in processing temporal dependence makes it an ideal selection in data association. When factors such as scene congestion and weak illumination cause detection failure especially long intervals, association is often very difficult and eventually leads to tracking failure. To solve this problem, Dual Scale Matching model (DSM) containing a Motion Trend Match Network (MTMNet) and an Appearance History Memory Network (AHMNet) is proposed. DSM is a long-term optimization method based on multiple hypothesis tracking. MTMNet aims to learn a similarity metric matching function between tracklets leveraging the motion feature. AHMNet is designed to provide optimal pruning strategies leveraging long period appearance feature. Our method is effective on MOT17 benchmark and it shows that we achieve considerable competitive results with current state-of-the-art trackers.

AB - Multi-object tracking can be characterized as a data association problem. The advantage of RNN in processing temporal dependence makes it an ideal selection in data association. When factors such as scene congestion and weak illumination cause detection failure especially long intervals, association is often very difficult and eventually leads to tracking failure. To solve this problem, Dual Scale Matching model (DSM) containing a Motion Trend Match Network (MTMNet) and an Appearance History Memory Network (AHMNet) is proposed. DSM is a long-term optimization method based on multiple hypothesis tracking. MTMNet aims to learn a similarity metric matching function between tracklets leveraging the motion feature. AHMNet is designed to provide optimal pruning strategies leveraging long period appearance feature. Our method is effective on MOT17 benchmark and it shows that we achieve considerable competitive results with current state-of-the-art trackers.

KW - Appearance History Memory Net

KW - Data association

KW - Dual Scale Matching

KW - Motion Trend Match Net

KW - Multi-object tracking

UR - http://www.scopus.com/inward/record.url?scp=85091508327&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-59016-1_54

DO - 10.1007/978-3-030-59016-1_54

M3 - Conference contribution

AN - SCOPUS:85091508327

SN - 9783030590154

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 653

EP - 665

BT - Wireless Algorithms, Systems, and Applications - 15th International Conference, WASA 2020, Proceedings

A2 - Yu, Dongxiao

A2 - Dressler, Falko

A2 - Yu, Jiguo

PB - Springer Science and Business Media Deutschland GmbH

T2 - 15th International Conference on Wireless Algorithms, Systems, and Applications, WASA 2020

Y2 - 13 September 2020 through 15 September 2020

ER -

Ye Z, Wu Y, Wang S, Zhang Y, Chen Y, Ke W et al. A Dual Scale Matching Model for Long-Term Association. In Yu D, Dressler F, Yu J, editors, Wireless Algorithms, Systems, and Applications - 15th International Conference, WASA 2020, Proceedings. Springer Science and Business Media Deutschland GmbH. 2020. p. 653-665. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). doi: 10.1007/978-3-030-59016-1_54