An Anti-UAV Long-Term Tracking Method with Hybrid Attention Mechanism and Hierarchical Discriminator

Feng Cheng, Zhibo Liang, Gaoliang Peng, Shaohui Liu, Sijue Li, Mengyu JiSensors 2022 · OpenReview

TL;DR

This paper introduces SiamAD, a novel anti-UAV long-term tracking system that combines a Siamese network with a hybrid attention mechanism and a hierarchical discriminator. By improving feature representation and reliably detecting tracking failures, SiamAD significantly enhances the accuracy and robustness of UAV tracking, outperforming existing state-of-the-art methods.

What Problem Does This Paper Address?

Unmanned Aerial Vehicles (UAVs) pose increasing security risks, making effective anti-UAV systems crucial. A core component of these systems is the ability to accurately and reliably track UAVs in real-world scenarios. However, tracking UAVs presents significant challenges: they are often small, appear in complex and dynamic environments with cluttered backgrounds, and can be subject to occlusion, out-of-view events, and rapid motion. Existing tracking methods, even advanced Siamese network-based trackers, struggle with these challenges, particularly in long-term tracking scenarios where targets can disappear and reappear, leading to drift and failure. This paper aims to address these limitations by developing a more robust and reliable long-term tracking solution specifically for anti-UAV applications.

Methods

The proposed SiamAD system is built upon a Siamese network architecture, which is a standard approach for visual tracking that treats tracking as a matching problem. The key innovations in SiamAD are:

  • Hybrid Attention Mechanism (MHAM): This module is integrated into the Siamese network's feature extraction backbone (a ResNet50). MHAM comprises three submodules: channel attention, spatial attention, and context attention. It aims to enhance feature representation by adaptively reweighting channel features, selectively highlighting informative spatial locations, and capturing semantic relationships between subregions. This helps to suppress background interference and focus on discriminative target features.
  • Hierarchical Discriminator (HDC): This component is crucial for long-term tracking. It consists of a Reliability Discriminator (RDC) and an Updating Discriminator (UDC). The RDC uses metrics like the RPN's classification score (Sp) and the Peak-to-Sidelobe Ratio (APCE) to assess the confidence of the current tracking result. If the confidence is low, the RDC triggers the redetection module. The UDC then uses these reliability assessments to decide when and how to update the target template.
  • Redetection Module: When tracking failures are detected by the HDC, a YOLOv5-based redetection network is employed to perform a global search and re-localize the target. This is critical for recovering from occlusions or out-of-view events.
  • Adaptive Template Updating Strategy: Instead of simply replacing the template with the latest detection, SiamAD uses an adaptive-weight strategy that considers historical information and the reliability of both tracking and redetection results. This helps to prevent the accumulation of noise and maintain a robust template.

Datasets

  • Anti-UAV Dataset: This dataset, introduced at CVPR 2020, serves as the primary benchmark. It contains over 300 RGB and infrared video pairs with more than 580k bounding boxes, covering diverse UAV types and challenging scenarios like out-of-view, occlusion, fast motion, scale variation, low illumination, thermal crossover, and low resolution.
  • Homemade Test Set: The authors also created a small, homemade test set using real UAV videos to evaluate the algorithm in actual scenarios.

Key Insights and Ideas

  • Attention for Robustness: The hybrid attention mechanism is key to improving feature discriminability, allowing the tracker to better distinguish targets from complex backgrounds and distractors, which is vital for UAV tracking.
  • Reliability is Paramount for Long-Term Tracking: Simply tracking is not enough; knowing when tracking has failed and having a reliable mechanism to recover is essential. The hierarchical discriminator provides this crucial reliability assessment.
  • YOLOv5 as an Effective Redetector: Leveraging a state-of-the-art object detector like YOLOv5 for redetection provides high accuracy and speed, enabling quick recovery from tracking failures.
  • Adaptive Template Updating Prevents Drift: A simple template update can lead to accumulated errors. The proposed adaptive strategy, which balances new information with historical data, is crucial for maintaining tracking accuracy over extended periods.

Key Results

  • State-of-the-Art Performance: SiamAD achieved the best results on the Anti-UAV benchmark, with an AUC of 67.7% and DP of 88.4%.
  • Significant Improvements: Compared to the strong baseline SiamRPN++, SiamAD showed improvements of 13.7% in success rate (AUC) and 16.5% in precision rate (DP). It also outperformed SiamRPN++LT by 7.3% and 9.0% in AUC and DP, respectively.
  • Robustness Across Challenges: SiamAD demonstrated superior performance in challenging scenarios like out-of-view (OV) and fast motion (FM), outperforming other trackers in these aspects.
  • Real-time Capability: The proposed method achieves a speed of 38.8 FPS, meeting real-time tracking requirements.
  • Ablation Studies Confirm Component Effectiveness: Experiments showed that each proposed component (hybrid attention, redetection, adaptive update) significantly contributed to the overall performance improvement.

Limitations and Future Work

  • Redetection Module Training: The redetection module is trained on a fixed dataset, and its performance can degrade if the distribution of the test set video differs significantly from the training set (as observed on the homemade test set).
  • Motion Information: The current framework does not explicitly consider UAV motion information, which could further improve robustness.
  • LSS-Target Tracking: The paper notes that tracking low-altitude, slow-speed, small targets (LSS-target) is an area that has received less attention and could be a focus for future optimization.

Future Work Suggestions:

  • Refine the search strategy of the redetection module and fuse it with the detection and tracking networks for parameter sharing and increased efficiency.
  • Develop a dynamic model incorporating spatiotemporal correlation to enhance tracking robustness.
  • Optimize the network for tracking low-altitude, slow-speed, and small targets.

Summary updated

Abstract

To prevent unmanned aerial vehicles (UAVs) from threatening public security, anti-UAV object tracking has become a critical issue in industrial and military applications. However, tracking UAV objects stably is still a challenging issue because the scenarios are complicated and the targets are generally small. In this article, a novel long-term tracking architecture composed of a Siamese network and re-detection (SiamAD) is proposed to efficiently locate UAV targets in diverse surroundings. Specifically, a new hybrid attention mechanism module is exploited to conduct more discriminative feature representation and is incorporated into a Siamese network. At the same time, the attention-based Siamese network fuses multilevel features for accurately tracking the target. We further introduce a hierarchical discriminator for checking the reliability of targeting, and a discriminator-based redetection network is utilized for correcting tracking failures. To effectively catch up with the appearance changes of UAVs, a template updating strategy is developed in long-term tracking tasks. Our model surpasses many state-of-the-art models on the anti-UAV benchmark. In particular, the proposed method can achieve 13.7% and 16.5% improvements in success rate and precision rate, respectively, compared with the strong baseline SiamRPN++.
BibTeX
@inproceedings{cheng2022anti,
  author = {Feng Cheng and Zhibo Liang and Gaoliang Peng and Shaohui Liu and Sijue Li and Mengyu Ji},
  title = {{An Anti-UAV Long-Term Tracking Method with Hybrid Attention Mechanism and Hierarchical Discriminator}},
  booktitle = {Sensors},
  year = {2022},
  url = {https://openreview.net/forum?id=89VfDqHfYJvO},
}
Figure