Real-Time Simultaneous Pose and Shape Estimation for Articulated Objects Using a Single Depth Camera.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Accurate pose and shape estimation of articulated objects is crucial for applications like human-computer interaction and robotics.
• Existing methods often struggle with fast, complex motions and require explicit point correspondences, limiting their robustness.
• Developing methods that are less sensitive to local minima and do not require extensive calibration is an ongoing challenge.

Purpose of the Study:

• To present a novel real-time algorithm for simultaneous pose and shape estimation of articulated objects.
• To develop a robust method that handles fast and complex motions effectively.
• To improve tracking accuracy through personalized shape modeling.

Main Methods:

• Embedding an articulated deformation model with exponential-maps-based parametrization into a Gaussian Mixture Model (GMM).
• Utilizing a probabilistic measurement model within the GMM to avoid explicit point correspondences.
• Implementing a novel shape adaptation algorithm that captures subject shape during dynamic pose estimation.
• Employing either a mesh model or a sphere-set model as a template within a unified framework.

Main Results:

• The proposed algorithm significantly outperforms state-of-the-art pose estimation methods, especially in challenging motion scenarios.
• Shape estimation achieves comparable accuracy to existing methods without requiring statistical shape models or extra calibration.
• The algorithm demonstrates high efficiency, achieving up to 60 frames per second on a middle-range GPU.

Conclusions:

• The novel algorithm provides accurate and robust simultaneous pose and shape estimation for articulated objects.
• The probabilistic GMM-based approach enhances tracking performance and reduces sensitivity to local minima.
• The GPU implementation enables real-time performance, making it suitable for practical applications.