UMOT: A unified framework for long- and short-term association for multi-object tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Multi-object tracking (MOT) faces challenges like trajectory breakage and identity switches in complex environments.
• Existing Transformer-based methods struggle with long-term dependency modeling and target recovery.

Purpose of the Study:

• To propose a unified framework, UMOT, addressing limitations in short-term motion prediction and long-term trajectory recovery for MOT.
• To enhance target recovery and identity preservation in challenging tracking scenarios.

Main Methods:

• UMOT integrates a pre-trained YOLOX detector and MOTR-ConvNext network for short-term correlation.
• Dynamically updated track and detect queries optimize short-term motion prediction.
• Track Query Memory Module (TQMM) and Historical Backtracking Module manage historical data and re-associate lost targets.

Main Results:

• UMOT demonstrates significant improvements in HOTA and IDF1 metrics on DanceTrack and MOT17 datasets.
• The framework shows robustness and effectiveness in scenarios with complex occlusions and long-term dependencies.

Conclusions:

• UMOT effectively alleviates the conflict between short-term motion prediction and long-term trajectory recovery.
• The proposed method offers a robust solution for multi-object tracking, particularly in challenging real-world conditions.