.png){kind=icon}
Return
GDOTrack: Multi-object tracking algorithm based on feature decoupling and appearance aggregation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patcog.2026.114444.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Joint Detection and Embedding (JDE) paradigm has been widely adopted for online multi-object tracking. However, detection and re-identification (ReID) tasks share a unified feature space in the JDE paradigm, which inherently leads to feature conflicts and limits tracking performance. To address this issue, we propose a multi-object tracking algorithm based on feature decoupling and appearance aggregation (GDOTrack). Firstly, a Global-Local Attention feature decoupling Network (GLANet) is proposed, which jointly captures local and global information for the detection and ReID branches to improve task-specific representation learning. Secondly, we introduce a Direction-Aware Appearance Feature Aggregation Network (DAAFAN) to aggregate multi-scale features along horizontal and vertical directions, enhancing the discriminability of object representation under scale variation and similar appearance. Finally, an Occlusion-aware Recovery IoU (ORIoU) method is designed to dynamically expand the bounding box based on the occlusion level, and we further build a multi-level association strategy (ORMA) to improve association robustness in occlusion scenarios. Experimental results on the MOT16, MOT17, and MOT20 benchmarks demonstrate that the proposed GDOTrack achieves state-of-the-art tracking performance.
Keywords:
Multi-object tracking
Global-local attention
Multi-scale appearance feature aggregation
Occlusion-aware IoU
Journal
IF:
7.6
Papers:
1.3W
Citations:
4.5W
Organization
Cited Papers
No cited papers available