ViTT: Vision Transformer Tracker.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Multi-object tracking (MOT) is vital for multi-unmanned aerial vehicle (Multi-UAV) systems.
• Traditional methods struggle with occlusion and complex environments.
• Transformers, successful in NLP, show promise in computer vision tasks.

Purpose of the Study:

• To propose a novel transformer-based model for multi-object tracking.
• To leverage the global context modeling of transformers for improved tracking accuracy.
• To address challenges in occlusion and complex scenarios within MOT.

Main Methods:

• Developed the Vision Transformer Tracker (ViTT), utilizing a transformer encoder as the backbone.
• Input images directly into the transformer architecture.
• Employed multi-task learning for simultaneous object localization and appearance embedding extraction.

Main Results:

• Achieved 65.7% MOTA on the MOT16 dataset.
• Demonstrated superior performance compared to other typical multi-object trackers.
• Showcased the effectiveness of transformer networks in complex computer vision tasks.

Conclusions:

• Transformer-based networks are highly effective for complex computer vision tasks like MOT.
• ViTT offers a promising direction for applying pure transformers in multi-object tracking.
• The model paves the way for advanced tracking capabilities in Multi-UAV applications.