Tracking Tiny Drones against Clutter: Large-Scale Infrared Benchmark with Motion-Centric Adaptive Algorithm

Jiahao Zhang, Zongli Jiang, Jinli Zhang, Yixin Wei, Liang Li, Yizheng Wang, Gang WangICCV 2025

TL;DR

This paper introduces MCATrack, a novel algorithm for tracking small drones in infrared videos, which are notoriously difficult to detect against complex backgrounds. They also present TDTIV, a large-scale infrared dataset specifically designed to challenge existing drone tracking methods. MCATrack achieves state-of-the-art performance by leveraging bio-inspired motion enhancement and a dynamic cross-guided module for adaptive target representation.

What Problem Does This Paper Address?

Tracking small drones, especially in infrared imagery, is a significant challenge for several reasons. Drones are often tiny (<20x20 pixels) and can blend seamlessly into complex, cluttered backgrounds like forests or urban environments. Existing drone tracking datasets often feature larger targets or simpler backgrounds, failing to adequately represent the difficulties of real-world scenarios. Furthermore, many current tracking algorithms rely heavily on appearance matching, which breaks down when drones change pose, are occluded, or experience drastic lighting variations. This makes robust, long-term tracking of these small, dynamic targets a critical open problem for applications like security and air traffic management.

Methods

The core of MCATrack is its Motion-Centric Adaptive Tracking (MCA-Track) approach, which combines two key novel modules:

  • Bio-inspired Magno-motion Enhancement Module: Inspired by the biological magnocellular pathway in vision, this module uses motion responses to amplify the signal-to-noise ratio of tiny targets. It effectively suppresses background clutter by focusing on spatio-temporal motion patterns, enhancing the contrast of the target. This is achieved through a recursive process that leverages temporal coherence across frames.
  • Dynamic Cross-Guided Module: This module addresses challenges like pose variations and occlusions. It operates in three stages:
  • Dynamic Target Cross-Guidance: A Siamese-branching Region Proposal Network (RPN) models the relationship between initial and updated target features to generate high-quality candidate regions.
  • Enhanced Feature Confidence Generation: This refines the proposed regions by integrating the latest target information with initial features, improving confidence accuracy. Unlike traditional methods that solely rely on the initial template, MCATrack uses updated target information for more precise box generation and refinement.
  • Dynamic Template Update: A strategy is employed to update the target's dynamic template in real-time, ensuring the tracker adapts to changing target appearances.

Datasets

  • TDTIV (Tiny Drone Tracking in Infrared Videos): This is a newly introduced, large-scale benchmark dataset. It comprises 290k frames with 280k manually annotated bounding boxes, specifically focusing on tiny drones in diverse and complex infrared environments (forests, buildings, clouds, mountains). The paper highlights that TDTIV's targets are generally smaller than those in existing datasets like Anti-UAV410, making it a more challenging benchmark.
  • Anti-UAV410: This is a well-recognized dataset used for evaluating the proposed method alongside TDTIV.

Key Insights and Ideas

  • Motion is Key for Tiny Targets: The paper strongly emphasizes that for tracking small, often featureless targets like drones, motion information is as crucial, if not more so, than appearance. Relying solely on appearance matching leads to failures in challenging scenarios.
  • Bio-Inspiration for Enhanced Contrast: Mimicking biological visual pathways (specifically the magnocellular system) provides an effective way to enhance the target's contrast against complex backgrounds, a critical step for detecting and tracking tiny objects.
  • Adaptive Target Representation is Crucial: The dynamic nature of drone tracking (pose changes, lighting, occlusions) necessitates an adaptive representation of the target. The Dynamic Cross-Guided module's ability to integrate updated target information is vital for maintaining tracking accuracy over time.
  • Dataset Design Matters: The creation of TDTIV underscores the need for specialized datasets that accurately reflect real-world challenges, particularly for emerging applications like drone monitoring.

Key Results

  • State-of-the-Art Performance: MCATrack achieves state-of-the-art results on both the proposed TDTIV dataset and the Anti-UAV410 dataset.
  • TDTIV Performance: On TDTIV, MCATrack achieved a State Accuracy (SA) of 54.97%, outperforming the previous best tracker, SiamDT, by over 5%. Precision and success plots show significant gains of 6.7% and 5.1% respectively compared to SiamDT.
  • Anti-UAV410 Performance: On Anti-UAV410, MCATrack ranked first, showing an improvement of over 10% compared to most competing algorithms and 1% over SiamDT. Precision and success plots indicated improvements of 3% and 1% respectively over SiamDT.
  • Ablation Study Insights:
    • The Magno-motion module alone contributed a significant 7.13% improvement in performance on TDTIV.
  • The Feature Enhancement module added a further 0.62% improvement on TDTIV.

Limitations and Future Work

  • Limitations: While the paper doesn't explicitly detail limitations, the inherent difficulty of tracking extremely small, fast-moving, and potentially occluded targets in infrared remains a challenge. The reliance on specific bio-inspired mechanisms might also have computational trade-offs not fully explored.
  • Future Work: The authors aim to explore more effective ways to integrate motion and temporal information to further enhance long-term tracking performance. This suggests continued research into refining the spatio-temporal understanding of target dynamics.

Summary updated

BibTeX
@inproceedings{zhang2025tracking,
  author = {Jiahao Zhang and Zongli Jiang and Jinli Zhang and Yixin Wei and Liang Li and Yizheng Wang and Gang Wang},
  title = {{Tracking Tiny Drones against Clutter: Large-Scale Infrared Benchmark with Motion-Centric Adaptive Algorithm}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2025},
}
Figure