A Multi-Object Tracking Method with an Unscented Kalman Filter on a Lie Group Manifold.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Multi-object tracking (MOT) is crucial but faces challenges with homogeneous appearance, heterogeneous motion, and heavy occlusion.
• Existing methods suffer from missed associations and false predictions due to simplified motion models and weak appearance representations.

Purpose of the Study:

• To propose a lightweight, generic, and appearance-independent MOT method.
• To address limitations in current MOT approaches for complex tracking scenarios.

Main Methods:

• Introduced LUKF-Track, a novel MOT method utilizing an unscented Kalman filter (UKF) on a Lie group.
• Employs a motion model for object state propagation and prediction formulated via UKF on the Lie group.
• Incorporates detection boxes across all score ranges for robust data association.

Main Results:

• Achieved state-of-the-art performance on the MOT17, MOT20, and DanceTrack benchmarks.
• Demonstrated effectiveness in scenarios with highly nonlinear motion and severe occlusions.
• Showcased improved tracking accuracy compared to existing methods.

Conclusions:

• LUKF-Track offers a significant advancement in multi-object tracking, particularly for challenging environments.
• The proposed UKF on a Lie group approach enhances robustness against appearance variations and motion complexities.