CSMOT: Make One-Shot Multi-Object Tracking in Crowded Scenes Great Again.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current one-shot multi-object tracking (MOT) algorithms, while fast, struggle with accuracy in crowded environments.
• Existing methods face challenges in precise object localization and distinguishing similar appearances, leading to identity switches.

Purpose of the Study:

• To develop a more robust multi-object tracking (MOT) algorithm, CSMOT, to address limitations in crowded scenes.
• To improve both the detection and re-identification components of MOT systems for enhanced stability and accuracy.

Main Methods:

• Incorporated a coordinate attention module into an encoder-decoder network to improve object detection capabilities.
• Introduced an angle-center loss function to enhance the discriminative power of re-identification features.
• Implemented a refined data association mechanism that considers all detections, not just high-scoring ones.

Main Results:

• CSMOT demonstrates excellent tracking performance on public datasets, particularly excelling in crowded scenarios.
• The proposed algorithm achieved significant reductions in identity switches: 11.8% on MOT16 and 33.8% on MOT17 compared to the baseline.
• CSMOT balances detection and re-identification tasks effectively through redesigned feature dimensions.

Conclusions:

• CSMOT offers a more robust and accurate solution for one-shot multi-object tracking, especially in challenging crowded scenes.
• The combination of improved detection, discriminative re-identification, and comprehensive data association leads to superior tracking performance.