A Collaborative Learning Tracking Network for Remote Sensing Videos.

Area of Science:

• Computer Vision
• Remote Sensing Technology
• Artificial Intelligence

Background:

• Remote sensing tracking is crucial but challenging in complex scenes.
• Existing methods struggle with small objects and background clutter.

Purpose of the Study:

• To develop an accurate and robust tracking network for remote sensing videos.
• To address challenges in feature extraction, object-background distinction, and interference handling.

Main Methods:

• Proposed a collaborative learning tracking network with three key modules: CRFPF, DSCA, and GCRT.
• CRFPF module: Extracts and fuses hierarchical features adaptively for small objects.
• DSCA module: Enhances object saliency and bounding box regression using spatial-channel co-attention.
• GCRT strategy: Mitigates interference from similar objects via motion trajectory analysis.

Main Results:

• The proposed network demonstrates feasibility and effectiveness on multiple datasets.
• Experimental results validate the performance improvements in complex remote sensing scenarios.
• The method successfully enhances object saliency and refines bounding box predictions.

Conclusions:

• The developed tracking network offers a significant advancement in remote sensing video analysis.
• The integrated modules effectively address key challenges in object detection and tracking.
• The approach provides a robust solution for real-world remote sensing applications.