TAPTrack: An Efficient Temporal-Aware Prompt Tracker for Infrared Anti-UAV Tracking

This is the official repository for TAPTrack: An Efficient Temporal-Aware Prompt Tracker for Infrared Anti-UAV Tracking (Accepted by IEEE Transactions on Geoscience and Remote Sensing (TGRS)).

📢 News

• [2026.04]: 🎉 Our paper has been accepted by TGRS
• [2026.04]: 📊 The raw tracking results on the AntiUAV410 dataset have been released.
• [Coming Soon]: The training code is currently being organized and will be released here soon.

---

📖 Abstract

With the widespread application of unmanned aerial vehicles (UAVs), efficient supervision and tracking have become urgent challenges. Tracking UAVs in infrared imaging is particularly difficult due to small size, sparse features, and significant susceptibility to camera motion and background clutter. Mainstream high-precision anti-UAV trackers adopt a global tracking paradigm; however, their inference speed often fails to meet the real-time demands of practical deployment. In contrast, the local tracking paradigm based on Transformer offers superior real-time performance. We found that enhancing the Search Factor (SF) can improve performance to some extent; however, this enhancement also introduces increased background noise. To address these issues, we propose an efficient local tracker, TAPTrack, specifically designed for infrared anti-UAV tracking. This method expands the SF while utilizing tracking results from previous frames as temporal prompts to guide the extraction of features from the search area. Additionally, a temporal prompt decoder (TPD) is developed to mitigate interference resulting from extensive searches. Furthermore, a training-friendly state discrimination mechanism is introduced, which combines the outputs of dual tracking heads with the Peak-to-Sidelobe Ratio (PSR) to determine the target’s status, thereby streamlining the training process. Finally, systematic evaluations were conducted on three infrared anti-UAV tracking benchmarks. Experimental results demonstrate that TAPTrack achieves superior performance among local trackers, showing a 2.1% improvement over the second-best method in the SA metric, while maintaining a real-time inference speed of 182 FPS.

---

📊 Tracking Results

We provide the raw tracking results of TAPTrack to facilitate comparison for future research.

Raw tracking results: Baidu, Password pspc

Note: You can use these .txt results directly in your evaluation toolkits to compare with our TAPTrack.

---

✒️ Citation

If you find our paper, code, or tracking results useful for your research, please consider citing:

@ARTICLE{cui2026taptrack,
  author={Cui, Xiaolong and Li, Xingxiu and Wu, Panlong and Liu, Xinan and He, Shan},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={TAPTrack: An Efficient Temporal-Aware Prompt Tracker for Infrared Anti-UAV Tracking}, 
  year={2026},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TGRS.2026.3682081}}

Acknowledgments

Thanks for the OSTrack and FocusTrack library, which helps us to quickly implement our ideas.