Multi-object tracking faces a major challenge in handling the variations of tracked targets within complex scenes. In existing transformer-based tracking methods, typically each tracked target is only associated with one track query. However, trajectories in crowded scenes often experience varying levels of occlusion, making the association brittle for using a single track query to identify the tracked target. Therefore, we argue that relying on a single track query to track a target in complex scenes is inadequate. In this paper, we introduce TGFormer, with the core idea of designing a Track Query Group for each tracked target. Each group encompasses track queries that handle the same tracked target across different levels of occlusion scenes. To achieve long-term robust association, we propose a novel updater that integrates temporal memories and occlusion-aware features to update the Track Query Group, ensuring the tracked target can be consistently captured in complex scenes. Additionally, we introduce a Position Predictor that allows TGFormer to forecast motion trends, helping the model accurately locate moving tracklets. Experimental results show that our method achieves competitive performance on the MOT Challenge and DanceTrack datasets.

Recent online Transformer-based multi-object tracking methods achieve end-to-end optimization by jointly performing detection and association. However, these trackers apply query augmentations uniformly to detect queries and track queries during training, which may limit their ability to fully exploit distinct features. The different augmentations serve distinct purposes and may have conflicting effects on detection and association. Moreover, the detect queries and the track queries with augmentations contain different semantic information. Jointly feeding these queries into self-attention modules for feature interaction may lead to suppression between the two types of queries. In this article, we present a novel Transformer-based end-to-end model, DQFormer, which mitigates conflicts and effectively learns task-specific features through the proposed Decoupled Query Augmentation (DQA) strategy. DQA categorizes query augmentations into separate detection and association branches, generating more discriminative queries tailored specifically for detection and association. To align with DQA, we decompose the self-attention module into a Dual Cross-Attention module. Furthermore, a Targeted Label Assignment strategy is applied to the augmented queries in each cross-attention module, helping the model to learn distinct features for tracking. DQFormer achieves competitive detection and association results across the MOT17, MOT20, and DanceTrack datasets.