Preformer MOT: A transformer-based approach for multi-object tracking with global trajectory prediction.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Multi-Object Tracking (MOT) methods often struggle with predicting future object trajectories.
• Existing approaches underutilize temporal information for robust trajectory prediction.
• Challenges include occlusions, overlapping objects, and nonlinear motion.

Purpose of the Study:

• To introduce Preformer MOT, a novel transformer-based method for precise nonlinear trajectory prediction.
• To enhance multi-object tracking by incorporating future frame information.
• To address the underutilization of temporal correlations in MOT.

Main Methods:

• Utilizes a transformer-based architecture (Preformer MOT).
• Combines a novel motion estimation technique with trajectory prediction and Kalman filtering.
• Employs a self-supervised trajectory prediction model for future position estimation.
• Incorporates multi-frame prediction to handle trajectory disruptions.

Main Results:

• Preformer MOT significantly enhances the precision of nonlinear trajectory predictions.
• The method demonstrates superior performance on pedestrian datasets (DanceTrack, MOT17).
• Achieves exceptional results in complex marine environments, showing adaptability.

Conclusions:

• Preformer MOT offers a comprehensive solution for multi-object tracking with extensive temporal correlations.
• The approach effectively predicts future trajectories, improving continuity during disruptions.
• Demonstrates broad applicability across diverse and challenging environments.