SuperTrack: Semantic Unified Pseudo-segmentation and Enhanced Representation for UAV Tracking

This is the official code of the paper SUPER-Track: Semantic Unified Pseudo-segmentation and Enhanced Representation for UAV Tracking.

Robust UAV tracking remains challenging in cluttered scenes due to severe appearance ambiguity, frequent occlusion, and strong motion interference, especially in multi-UAV scenarios. To address these issues, we propose SuperTrack, a two-stage UAV tracking framework composed of a pseudo-segmentation-enhanced detector and a semantic-assisted trajectory association module. In the detection stage, we introduce an instance semantic pseudo-segmentation strategy to improve foreground-background separability for small UAVs, a spatio-temporal consistency regularization to preserve temporal coherence under noise and occlusion, and a background hard negative mining scheme to suppress confusing background responses. In the association stage, instead of relying solely on frame-level appearance cues, we summarize recent trajectory statistics within temporal windows and convert them into template-based textual prompts, which are encoded by a frozen text encoder into compact semantic embeddings for association refinement. This design complements geometry-based matching with a structured semantic context extracted from short-term trajectory statistics, providing a stable, structured semantic representation of trajectory cues. Extensive experiments on Track 3, Anti-UAV410, LaSOT, LaSOT$_{ext}$, GOT-10K, and TrackingNet demonstrate that SuperTrack achieves competitive performance across both UAV-specific and generic tracking benchmarks.

The structure of the UAV detection part is shown as subfigure:

Correspondingly, the structure of UAV target tracking is shown in the following figure:

We can easily start training the SUPERTrack model by:

python train.py

Then run the following code to train the tracker:

python csptracker.py

Calculate the MOTA value using the detection code officially released by CVPR. First, enter folder MultiUAV_Baseline_code_and_submissi, then run:

python tool/2_compute_MOTA.py

And you also can test the trained model by run:

python test.py

The visualization of SUPERTrack on AntiUAV410 datatset is shown in the following figure:

Acknowledgement

Thanks for the Dist-Tracker and YOLOv13 library, which helps us to quickly implement our ideas.

Citation

If our work is useful for your research, please consider citing.