Fusing Appearance and Spatio-Temporal Models for Person Re-Identification and Tracking

Area of Science:

• Computer Vision
• Artificial Intelligence
• Pattern Recognition

Background:

• Accurate person localization is crucial for many computer vision tasks.
• Existing methods often rely on either appearance-based re-identification or spatio-temporal tracking, with limitations in each.
• Integrating both approaches can potentially overcome individual method weaknesses.

Purpose of the Study:

• To develop and evaluate a novel model fusion approach for enhancing person re-identification and tracking accuracy.
• To combine the strengths of appearance-based re-identification and spatio-temporal tracking models.
• To improve the overall accuracy of determining identity classes for detected people.

Main Methods:

• A Sequential k-Means algorithm for appearance-based re-identification.
• A Kalman filter for spatio-temporal tracking.
• A linear weighting approach to fuse model outputs, with adaptive weight adjustments using a decay function and rule-based system.

Main Results:

• Preliminary experiments demonstrate that fusing appearance and spatio-temporal models significantly improves classification accuracy.
• The proposed fusion method shows enhanced performance compared to individual models.
• Results were validated using two different person detection algorithms on an indoor dataset.

Conclusions:

• Model fusion of appearance and spatio-temporal information offers a significant advantage for person re-identification and tracking.
• The adaptive weighting strategy effectively balances the contributions of each model.
• This integrated approach represents a promising advancement in accurate person localization for computer vision applications.