Studentized dynamical system for robust object tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Dynamical systems (DS) are used for temporal sequence modeling.
• Probabilistic principal component analysis (PPCA) models object appearances but is sensitive to outliers.
• Gaussian distributions in PPCA have light tails, limiting robustness in real-world data.

Purpose of the Study:

• To develop a more robust target tracking system using dynamical systems.
• To address the limitations of Gaussian-based observation models in handling outliers.
• To improve tracking performance in the presence of significant noise.

Main Methods:

• Proposed a studentized dynamical system (SDS) by augmenting traditional DS with auxiliary latent variables.
• Adjusted the shape of the observation distribution using these latent variables.
• Employed heavy-tailed distributions, specifically inspired by the Student's t-distribution, for enhanced robustness.

Main Results:

• The proposed SDS demonstrates improved capability in handling outlier noise.
• Achieved better tracking performance compared to traditional DS with Gaussian-based models.
• The use of auxiliary latent variables effectively created heavier-tailed observation distributions.

Conclusions:

• Studentized dynamical systems offer a robust approach to target tracking.
• The SDS framework provides superior performance in noisy environments with outliers.
• This method advances appearance modeling for visual tracking applications.