Transformer-Based Multiple-Object Tracking via Anchor-Based-Query and Template Matching.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Multiple object tracking (MOT) is crucial for intelligent video analysis, with joint detection and tracking (JDT) methods offering a unified approach.
• Current JDT methods face limitations in training speed and data association accuracy, hindering their overall performance.

Purpose of the Study:

• To enhance the efficiency and effectiveness of JDT methods for MOT.
• To introduce novel modules for faster training convergence and improved data association in MOT.

Main Methods:

• Proposed an anchor-based query (ABQ) module that integrates anchor box coordinates into learnable queries, providing explicit spatial priors for focused feature learning.
• Developed a template matching (TM) module to enable JDT methods to associate current detections with historical trajectory features.
• Introduced ABQ-Track, a novel transformer-based MOT method incorporating both ABQ and TM modules.

Main Results:

• The ABQ module significantly accelerates training convergence, reducing training epochs from 150 to 50 compared to baseline methods.
• The TM module enhances the data association capabilities by leveraging historical feature information.
• ABQ-Track demonstrated superior performance over existing JDT methods like TransTrack in extensive experiments.

Conclusions:

• The proposed ABQ and TM modules effectively address the limitations of existing JDT methods in MOT.
• ABQ-Track offers a more efficient and accurate solution for multiple object tracking tasks.
• The integration of spatial priors and historical feature matching represents a significant advancement in JDT methodology.