Dynamic Human Body Modeling Using a Single RGB Camera.

Area of Science:

• Computer Vision
• 3D Human Body Modeling
• Computer Graphics

Background:

• Accurate 3D human body modeling is crucial for various applications, including virtual reality and animation.
• Existing methods often require multiple cameras, depth sensors, or manual user input, limiting their accessibility.
• Dynamic modeling from monocular RGB data presents challenges due to inherent depth ambiguity.

Purpose of the Study:

• To develop an automatic pipeline for generating personalized parametric 3D human body models from single RGB camera input.
• To leverage human motion for incremental and automatic model construction.
• To achieve accuracy comparable to depth-based methods without specialized hardware.

Main Methods:

• Utilizing coarse 2D and 3D pose estimation from image sequences.
• Performing kinematic classification of human body parts to refine poses and reconstruct body parts.
• Driving a general parametric human template to fit reconstructed body parts and calculating non-rigid deformation.

Main Results:

• The proposed method automatically and incrementally models dynamic 3D human bodies using only an RGB camera.
• The system refines poses and generates personalized parametric models by fitting a template to reconstructed body parts.
• Shape estimation accuracy is comparable to methods using depth cameras.

Conclusions:

• The developed pipeline enables automatic and personalized 3D human body modeling from widely available RGB cameras.
• The approach effectively utilizes human motion for robust model generation.
• This method significantly enhances the feasibility of advanced 3D human modeling on standard devices like smartphones.