Modeling of Multiple Spatial-Temporal Relations for Robust Visual Object Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• One-stream trackers using Transformer architectures excel at parallel feature extraction and relation modeling.
• However, these trackers struggle in complex scenarios due to overlooking crucial tracking cues beyond the template.

Purpose of the Study:

• To propose a novel multi-cue single-stream tracker (MCTrack) for improved visual object tracking.
• To seamlessly integrate template information, historical trajectory, historical frame, and search region for synchronized feature extraction and relation modeling.

Main Methods:

• MCTrack employs two encoder types to convert diverse tracking cues into tokens for a Transformer architecture.
• A novel adaptive update mechanism with thresholding and local multi-peak components refines temporal and spatial cue distillation.

Main Results:

• MCTrack demonstrates leading performance on mainstream benchmark datasets.
• Achieved a 2.0% improvement in the AO metric on GOT-10k compared to the advanced SeqTrack.

Conclusions:

• The proposed MCTrack effectively addresses limitations of one-stream trackers in complex visual tracking tasks.
• The integration of multiple cues and an adaptive update mechanism significantly boosts tracking accuracy and robustness.