Motion Capture With Ellipsoidal Skeleton Using Multiple Depth Cameras.

Area of Science:

• Computer Vision
• Robotics
• Human-Computer Interaction

Background:

• Marker-less motion capture is crucial for applications like virtual reality and animation.
• Existing depth camera-based methods often struggle with complex human motions and geometric detail.
• Accurate skeletal fitting to point cloud data remains a significant challenge.

Purpose of the Study:

• To introduce a novel motion capturing framework utilizing ellipsoid-based skeletons.
• To enhance geometric detail capture using spherical harmonics encoded displacement and normal functions.
• To develop a robust and efficient marker-less motion tracking system for depth camera data.

Main Methods:

• Minimizing fitting error between an ellipsoid-based skeleton and input point cloud data.
• Employing ellipsoids with spherical harmonics for detailed geometry and normal representation.
• Integrating a bone collision avoidance mechanism into the motion capturing process.
• Parallel implementation using CUDA on Graphics Processing Units (GPUs) for speed.

Main Results:

• The framework achieves errors within a reasonable range compared to ground truth on the MHAD dataset.
• Successfully captures challenging motions that fail with Microsoft Kinect SDK and are not addressed by prior work.
• Demonstrates advantages in robustness and input data modality compared to state-of-the-art marker-less depth camera methods.
• Exhibits fast running speed due to parallel GPU implementation.

Conclusions:

• The proposed ellipsoid-based framework offers a robust and accurate solution for marker-less motion capture using depth cameras.
• The integration of spherical harmonics significantly improves the capture of geometric details.
• The method shows superior performance on complex motions, expanding the capabilities of current motion tracking systems.