Track creation and deletion framework for long-term online multiface tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing visual tracking methods often focus on feature extraction or cue fusion.
• Effective track management, crucial for multi-object tracking, is an under-addressed challenge.
• Deciding when to initiate or terminate tracking is complex due to detector limitations and model variability.

Purpose of the Study:

• To address the track management problem in multi-object tracking.
• To present a real-time online multiface tracking algorithm with improved track management.
• To enhance tracking robustness by effectively deciding target addition and removal.

Main Methods:

• Formulated tracking within a multi-object state-space Bayesian filtering framework.
• Employed Markov Chain Monte Carlo (MCMC) for solving the Bayesian filtering problem.
• Implemented an explicit probabilistic filtering step for track management decisions, utilizing face detections, likelihood measures, long-term observations, and track state characteristics.

Main Results:

• Demonstrated significant performance increases compared to traditional approaches.
• The proposed method outperformed traditional Markov Chain Monte Carlo (MCMC) and reversible-jump Markov Chain Monte Carlo (RJMCMC) methods.
• Achieved superior results on challenging datasets totaling over 9 hours of data.

Conclusions:

• The proposed track management strategy effectively handles difficulties arising from object detector deficiencies and observation model limitations.
• The integrated probabilistic filtering step provides a robust mechanism for managing object tracks in real-time.
• This approach offers a significant advancement in multi-object tracking applications, particularly for scenarios involving complex visual data.