MotionTrack: Learning motion predictor for multiple object tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Multi-object tracking (MOT) has advanced with detection and re-identification (ReID) but struggles with homogeneous appearance and heterogeneous motion.
• Current limitations stem from ReID feature discriminability and reliance on linear motion models.

Purpose of the Study:

• Introduce MotionTrack, a novel motion-based tracker.
• Enhance MOT accuracy by focusing on learnable motion prediction from trajectory data.

Main Methods:

• Developed a learnable motion predictor using only object trajectory information.
• Integrated token-level (self-attention) and channel-level (Dynamic MLP) motion features.
• Implemented a simple, online tracking approach.

Main Results:

• Achieved state-of-the-art performance on challenging datasets like Dancetrack and SportsMOT.
• Demonstrated superior accuracy in scenarios with complex object motion.
• Validated the effectiveness of the proposed motion prediction strategy.

Conclusions:

• MotionTrack offers a robust solution for MOT challenges, particularly in complex motion scenarios.
• The motion-based approach effectively models temporal dynamics for precise object prediction.
• This work advances MOT by leveraging sophisticated motion modeling beyond traditional methods.