Effective Multi-Object Tracking via Global Object Models and Object Constraint Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Multi-object tracking (MOT) faces challenges balancing accuracy and speed.
• Current MOT methods rely on object appearance and motion models, leading to high computational complexity.
• Reducing computational complexity is key to effective MOT.

Purpose of the Study:

• To propose global appearance and motion models for improved object discrimination.
• To introduce object constraint learning for enhanced tracking efficiency.
• To integrate global models and constraint learning into a confidence-based association method.

Main Methods:

• Learning a global appearance model via contrastive learning.
• Learning a global relation motion model using relative motion.
• Implementing object constraint learning, updating models only when discriminability constraints are violated.

Main Results:

• Achieved 64.5% Multi-Object Tracking Accuracy (MOTA) on the MOT16 test dataset.
• Reached a tracking speed of 6.54 Hz.
• Demonstrated simultaneous improvement in both tracking accuracy and speed compared to state-of-the-art methods.

Conclusions:

• Global models and object constraint learning effectively address the accuracy-speed trade-off in MOT.
• The proposed method offers a more efficient alternative to traditional per-frame online learning.
• This approach significantly advances the field of multi-object tracking.