Unsupervised Non-Rigid Human Point Cloud Registration Based on Deformation Field Fusion

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Data Processing

Background:

• Human point cloud registration is vital for computer vision and graphics.
• Existing methods struggle with joint hinges and limb occlusions in human point clouds.

Purpose of the Study:

• To propose an unsupervised non-rigid human point cloud registration method.
• To address challenges posed by joint hinges and limb occlusions.

Main Methods:

• A novel deformation field fusion approach is presented.
• It combines a deep dynamic link deformation field estimation module (using time series networks for rigid deformation conversion) and a probabilistic alignment deformation field estimation module (using Gaussian mixture models with constraints).

Main Results:

• The method demonstrates higher accuracy and robustness on public and private datasets.
• It effectively handles conditions with joint hinges and limb occlusions.

Conclusions:

• The proposed unsupervised non-rigid registration method successfully overcomes limitations of existing techniques.
• Deformation field fusion enhances sensitivity to both global and local features, improving human point cloud alignment.