High-Precision Multi-Object Tracking in Satellite Videos via Pixel-Wise Adaptive Feature Enhancement.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Multi-target tracking (MOT) in satellite imagery presents challenges due to small object scales and bird's-eye view perspectives.
• Objects like aircraft, ships, and vehicles often exhibit similar visual appearances, complicating tracking.
• Existing methods may struggle with efficiency and accuracy for these specific conditions.

Purpose of the Study:

• To develop an efficient and accurate end-to-end joint detection and tracking (JDT) method for satellite videos.
• To improve the learning of object representations for small-scale, bird's-eye view targets.
• To enhance tracking accuracy for objects with similar appearances.

Main Methods:

• A transformer-distillation-based end-to-end JDT method is proposed.
• A pixel-wise transformer-based feature distillation module is introduced for learning object representations.
• A temperature-controllable key feature learning objective function is utilized to improve the handling of similar-looking objects.
• Knowledge distillation from a precise teacher network and tracking head is employed within an end-to-end training framework.

Main Results:

• The proposed method demonstrates superior performance on three publicly available satellite video datasets.
• Over 90% overall tracking performance was achieved on the AIR-MOT dataset.
• The approach effectively improves tracking accuracy without compromising computational efficiency.

Conclusions:

• The transformer-distillation-based JDT method offers a significant advancement in multi-target tracking for satellite videos.
• The proposed techniques effectively address challenges related to object scale, viewpoint, and appearance similarity.
• This work provides a robust and efficient solution for satellite video analysis.