Multi-Object Tracking with Confidence-Based Trajectory Prediction Scheme.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Multi-Object Tracking (MOT) associates objects across video frames for stable trajectory maintenance.
• Current MOT methods often filter detections based on confidence scores, neglecting full utilization of detection results.
• Kalman Filter (KF) is widely used in MOT for sequential frame processing but struggles with noise in crowded scenes, leading to identity switches (IDS) and tracking failures.

Purpose of the Study:

• To investigate the limitations of existing trajectory prediction schemes in Multi-Object Tracking (MOT).
• To demonstrate that the Kalman Filter (KF) can achieve competitive results in video sequence processing with proper noise handling.
• To propose a novel confidence-based trajectory prediction scheme (ConfMOT) to improve MOT performance.

Main Methods:

• Proposed a confidence-based trajectory prediction scheme (ConfMOT) leveraging the Kalman Filter (KF).
• Utilized detection confidence scores (CS) to adjust KF noise during updates and predict object trajectories.
• Implemented a cost matrix (CM) for matching unreliable objects and removed lost trajectories based on CS.

Main Results:

• The ConfMOT tracker demonstrated superior performance compared to advanced competitors on mainstream datasets.
• The proposed method effectively handles noise in crowded scenes, reducing identity switches (IDS) and tracking failures.
• ConfMOT proved to be a simple yet efficient tracking solution.

Conclusions:

• The Kalman Filter (KF) remains a viable and competitive approach for Multi-Object Tracking (MOT) when noise is properly managed.
• The confidence-based trajectory prediction scheme (ConfMOT) offers a significant improvement in tracking accuracy and stability.
• ConfMOT provides an efficient and effective solution for Multi-Object Tracking challenges, particularly in complex environments.