SiamHSFT: A Siamese Network-Based Tracker with Hierarchical Sparse Fusion and Transformer for UAV Tracking

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Unmanned Aerial Vehicle (UAV) tracking faces challenges like low resolution, fast motion, and background interference due to platform limitations.
• Existing algorithms struggle to balance tracking performance and computational efficiency for UAV applications.

Purpose of the Study:

• To propose a novel Unmanned Aerial Vehicle (UAV) tracking algorithm, SiamHSFT, that achieves a balance between tracking robustness and real-time computation.
• To enhance feature extraction and target discrimination for improved tracking accuracy in complex UAV scenarios.

Main Methods:

• Developed SiamHSFT based on the Siamese network framework.
• Integrated CBAM attention and downward information interaction for feature enhancement, preserving information for small targets.
• Employed an interlaced sparse attention module focusing on spatial intervals to utilize global context effectively.
• Optimized the Transformer encoder with a modulation enhancement layer and triplet attention to improve inter-layer dependencies and target discrimination.

Main Results:

• SiamHSFT demonstrated excellent performance across diverse UAV tracking datasets (UAV123, UAV20L, UAV123@10fps, DTB70).
• Achieved superior performance in scenarios with fast motion and dynamic blurring.
• Maintained an average tracking speed of 126.7 fps, meeting real-time tracking requirements.

Conclusions:

• SiamHSFT effectively addresses the challenges of UAV target tracking by enhancing feature representation and attention mechanisms.
• The algorithm offers a robust and efficient solution for real-time tracking applications on Unmanned Aerial Vehicle platforms.