Details
| Original language | English |
|---|---|
| Pages (from-to) | 1619-1626 |
| Number of pages | 8 |
| Journal | International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences - ISPRS Archives |
| Volume | 48 |
| Issue number | 1 |
| Publication status | Published - 14 Dec 2023 |
| Event | ISPRS Geospatial Week 2023 - Kairo, Egypt Duration: 2 Sept 2023 → 7 Sept 2023 |
Abstract
Recent advancements in multi-object tracking (MOT) have heavily relied on object detection models, with attention-based models like DEtection TRansformer (DETR) demonstrating state-of-the-art capabilities. However, the utilization of attention-based detection models in tracking poses a limitation due to their large parameter count, necessitating substantial training data and powerful hardware for parameter estimation. Ignoring this limitation can lead to a loss of valuable temporal information, resulting in decreased tracking performance and increased identity (ID) switches. To address this challenge, we propose a novel framework that directly incorporates motion priors into the tracking attention layer, enabling an end-to-end solution. Our contributions include: I) a novel approach for integrating motion priors into attention-based multi-object tracking models, and II) a specific realisation of this approach using a Kalman filter with a constant velocity assumption as motion prior. Our method was evaluated on the Multi-Object Tracking dataset MOT17, initial results are reported in the paper. Compared to a baseline model without motion prior, we achieve a reduction in the number of ID switches with the new method.
Keywords
- Attention, Image Sequence Analysis, Motion Modelling, Pedestrian Tracking, Transformer
ASJC Scopus subject areas
- Computer Science(all)
- Information Systems
- Social Sciences(all)
- Geography, Planning and Development
Cite this
- Standard
- Harvard
- Apa
- Vancouver
- BibTeX
- RIS
In: International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences - ISPRS Archives, Vol. 48, No. 1, 14.12.2023, p. 1619-1626.
Research output: Contribution to journal › Conference article › Research › peer review
}
TY - JOUR
T1 - Integrating Motion Priors For End-To-End Attention-Based Multi-Object Tracking
AU - Ali, R.
AU - Mehltretter, M.
AU - Heipke, C.
N1 - Funding Information: This work was supported by the German Research Foundation (DFG) as a part of the Research Training Group i.c.sens [GRK2159].
PY - 2023/12/14
Y1 - 2023/12/14
N2 - Recent advancements in multi-object tracking (MOT) have heavily relied on object detection models, with attention-based models like DEtection TRansformer (DETR) demonstrating state-of-the-art capabilities. However, the utilization of attention-based detection models in tracking poses a limitation due to their large parameter count, necessitating substantial training data and powerful hardware for parameter estimation. Ignoring this limitation can lead to a loss of valuable temporal information, resulting in decreased tracking performance and increased identity (ID) switches. To address this challenge, we propose a novel framework that directly incorporates motion priors into the tracking attention layer, enabling an end-to-end solution. Our contributions include: I) a novel approach for integrating motion priors into attention-based multi-object tracking models, and II) a specific realisation of this approach using a Kalman filter with a constant velocity assumption as motion prior. Our method was evaluated on the Multi-Object Tracking dataset MOT17, initial results are reported in the paper. Compared to a baseline model without motion prior, we achieve a reduction in the number of ID switches with the new method.
AB - Recent advancements in multi-object tracking (MOT) have heavily relied on object detection models, with attention-based models like DEtection TRansformer (DETR) demonstrating state-of-the-art capabilities. However, the utilization of attention-based detection models in tracking poses a limitation due to their large parameter count, necessitating substantial training data and powerful hardware for parameter estimation. Ignoring this limitation can lead to a loss of valuable temporal information, resulting in decreased tracking performance and increased identity (ID) switches. To address this challenge, we propose a novel framework that directly incorporates motion priors into the tracking attention layer, enabling an end-to-end solution. Our contributions include: I) a novel approach for integrating motion priors into attention-based multi-object tracking models, and II) a specific realisation of this approach using a Kalman filter with a constant velocity assumption as motion prior. Our method was evaluated on the Multi-Object Tracking dataset MOT17, initial results are reported in the paper. Compared to a baseline model without motion prior, we achieve a reduction in the number of ID switches with the new method.
KW - Attention
KW - Image Sequence Analysis
KW - Motion Modelling
KW - Pedestrian Tracking
KW - Transformer
UR - http://www.scopus.com/inward/record.url?scp=85183301229&partnerID=8YFLogxK
U2 - 10.5194/isprs-archives-XLVIII-1-W2-2023-1619-2023
DO - 10.5194/isprs-archives-XLVIII-1-W2-2023-1619-2023
M3 - Conference article
AN - SCOPUS:85183301229
VL - 48
SP - 1619
EP - 1626
JO - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences - ISPRS Archives
JF - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences - ISPRS Archives
SN - 1682-1750
IS - 1
T2 - ISPRS Geospatial Week 2023
Y2 - 2 September 2023 through 7 September 2023
ER -