Yuhan Liu

Yuhan Liu

Associate Professor

Anti-UAV Tracking in Thermal Infrared

Published:

A Contrastive-Augmented Memory Network for Anti-UAV Tracking in TIR Videos
Z Wang, Y Hu, J Yang, G Zhou, F Liu, Y Liu
Published in Remote Sensing, December 2024
DOI

Project Overview

Challenge: Anti-UAV Tracking

The proliferation of consumer drones creates urgent need for detection and tracking systems.

TIR Video Challenges:

  • Thermal crossover: Target and background similar temperatures
  • Small targets: Distant drones appear as tiny blobs
  • Occlusions: Buildings, trees block view
  • Similar objects: Birds, kites cause confusion
  • Fast motion: Rapid maneuvering

Solution: CAM-Net

Contrastive-Augmented Memory Network:

  1. Memory network for long-term appearance modeling
  2. Contrastive learning for better discrimination
  3. Dynamic template update for adaptation

Key Innovations

1. Dual Memory Architecture

Short-term memory:

  • Recent appearance variations
  • Quick adaptation to changes

Long-term memory:

  • Historical appearance templates
  • Occlusion recovery capability

2. Contrastive Learning

Objective: Distinguish target from background and distractors

Strategy:

  • Positive pairs: Target templates from different times
  • Negative pairs: Background patches and similar objects
  • InfoNCE loss for discrimination

3. Dynamic Memory Update

Selective storage:

  • High-confidence detections → Store
  • Uncertain frames → Discard
  • Prevent drift and corruption

Method

Network Architecture

Query Frame
    ↓
Feature Extraction (CNN)
    ↓
Memory Retrieval (Attention)
├── Short-term memory matching
├── Long-term memory matching
└── Fusion
    ↓
Contrastive Enhancement
    ↓
Target Localization
    ↓
Memory Update (conditional)

Results

Anti-UAV Challenge benchmark:

  • Superior performance on public TIR dataset
  • Robust to occlusions (60%+ occlusion recovery)
  • Distinguishes drones from birds effectively

Applications

  • Airport security: Detect unauthorized drones near runways
  • Critical infrastructure: Power plants, government buildings
  • Public events: Concerts, sports events protection
  • Border surveillance: Monitor illegal activities

Citation

@article{wang2024contrastive,
  title={A Contrastive-Augmented Memory Network for Anti-UAV Tracking in TIR Videos},
  author={Wang, Z and Hu, Y and Yang, J and Zhou, G and Liu, F and Liu, Y},
  journal={Remote Sensing},
  volume={16},
  number={24},
  pages={4775},
  year={2024},
  publisher={MDPI}
}