SPT: Single Pedestrian Tracking Framework with Re-Identification-Based Learning Using the Siamese Model

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Pedestrian tracking is crucial for surveillance, robotics, and autonomous driving.
• Existing methods face challenges with occlusions, illumination changes, and appearance variations.

Purpose of the Study:

• To develop an improved single pedestrian tracking (SPT) framework.
• To enhance pedestrian re-identification accuracy and overall tracking robustness.

Main Methods:

• A tracking-by-detection paradigm combining deep learning and metric learning.
• Development of two compact metric learning models using Siamese architecture for re-identification.
• Integration of a robust re-identification model with a pedestrian detector for tracking.

Main Results:

• Re-identification models achieved accuracies of up to 96% on test datasets.
• The SPT tracker outperformed SOTA trackers in success rate (79.7%) and speed (18 FPS).
• Demonstrated effectiveness under various environmental challenges like illumination and occlusion.

Conclusions:

• The proposed SPT framework offers significant advancements in single pedestrian tracking.
• The novel re-identification models and integrated tracking approach enhance performance and robustness.