A 3D Relative-Motion Context Constraint-Based MAP Solution for Multiple-Object Tracking Problems.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Multi-object tracking (MOT) with moving monocular cameras is challenging due to reliance on detection accuracy.
• Occlusions and detection errors frequently cause track fragmentation and drift in traditional methods.

Purpose of the Study:

• To develop a unified framework for simultaneous multi-object tracking and camera motion estimation.
• To improve tracking robustness against occlusions and mis-detections by incorporating 3D object relations.

Main Methods:

• Formulated MOT and camera motion estimation as a joint maximum a posteriori (MAP) probability problem.
• Integrated a 3D relative-motion model within a sequential Bayesian framework.
• Employed reversible jump Markov chain Monte Carlo (RJMCMC) particle filtering for posterior estimation.

Main Results:

• The proposed method demonstrates superior performance across various evaluation metrics on benchmark datasets and real-world videos.
• Successfully addressed challenges of object occlusion and detection failures.
• Achieved robust tracking and accurate camera ego-motion estimation.

Conclusions:

• The unified framework effectively enhances multi-object tracking and camera motion estimation.
• The 3D relative-motion model is crucial for robustly handling occlusions and predicting object states.
• The RJMCMC particle filtering approach provides an effective solution for the complex posterior estimation problem.