Articulated and Generalized Gaussian Kernel Correlation for Human Pose Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Human pose estimation is crucial for human-computer interaction and activity recognition.
• Existing methods often struggle with complex articulations and subject-specific shapes.

Purpose of the Study:

• To propose a generalized and articulated Gaussian kernel correlation (GKC) framework for human pose estimation.
• To develop a robust tracking algorithm for full body and hand poses using depth data.

Main Methods:

• Derivation of a unified GKC representation generalizing sum of Gaussians (SoG) methods.
• Integration of a kinematic skeleton into a multivariate SoG template for articulated GKC (AGKC).
• Development of a sequential pose tracking algorithm with visibility, intersection, and pose continuity regularization.

Main Results:

• The proposed AGKC framework effectively models subject-specific shapes and estimates articulated poses.
• The sequential tracking algorithm demonstrates simplicity, effectiveness, and computational efficiency.
• Promising and competitive results achieved on benchmark depth datasets compared to state-of-the-art methods.

Conclusions:

• The articulated GKC framework offers a generalized and powerful approach to human pose estimation.
• The developed tracking algorithm provides an efficient and accurate solution for real-time applications.
• This work advances the state-of-the-art in depth-based human pose and hand tracking.