Tracking-by-Fusion via Gaussian Process Regression Extended to Transfer Learning

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object tracking is crucial in computer vision.
• Existing methods face challenges with appearance variations and computational efficiency.
• Transfer learning offers potential for improving tracking performance by leveraging auxiliary data.

Purpose of the Study:

• To develop a novel object tracking framework using Gaussian Processes (GPs) and particle filtering.
• To integrate Gaussian Process Regression (GPR) with transfer learning for enhanced appearance modeling.
• To fuse GPs with Correlation Filters (CFs) for improved sampling and tracking accuracy.

Main Methods:

• A novel Gaussian Processes (GPs)-based particle filter tracking framework is proposed.
• Gaussian Process Regression (GPR) is extended for transfer learning, utilizing auxiliary and target labeled samples, along with unlabeled data.
• A tracking-by-fusion strategy integrates a GPs component for appearance modeling and a Correlation Filters (CFs) component for efficient particle sampling.
• The CFs component benefits from the GPs component through re-weighted knowledge as latent variables.

Main Results:

• The proposed framework demonstrates superior performance on four benchmark datasets: OTB-2015, Temple-Color, and VOT2015/2016.
• Comparative analysis against baseline and state-of-the-art trackers validates the framework's effectiveness.
• The fusion of GPs and CFs, enabled by transfer learning, leads to significant improvements in object tracking accuracy and robustness.

Conclusions:

• The developed GPs-based particle filter framework offers a powerful and effective solution for object tracking.
• The integration of transfer learning with GPs and CFs enhances appearance modeling and sampling quality.
• The framework's superior performance highlights the potential of combining probabilistic methods with efficient filtering techniques for advanced computer vision tasks.