Kinect Posture Reconstruction Based on a Local Mixture of Gaussian Process Models.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Depth sensor hardware like Kinect enables interactive applications but faces challenges in accurate 3D human motion estimation.
• Noisy depth data and self-occlusion in user actions limit the precision of current posture recognition systems.

Purpose of the Study:

• To propose a novel real-time probabilistic framework for enhancing the accuracy of live-captured 3D human postures.
• To address the computational complexity of Gaussian Processes for large motion capture databases.

Main Methods:

• Utilized Gaussian Process models as a prior, incorporating positional data from depth sensors and marker-based systems.
• Integrated a temporal consistency term to minimize velocity variations between successive frames.
• Developed a local mixture of Gaussian Processes to reduce learning complexity and enhance prediction speed.

Main Results:

• The proposed framework significantly improves the accuracy of real-time 3D human posture estimation.
• Local mixture of Gaussian Processes drastically reduces learning time and increases prediction speed compared to standard Gaussian Processes.
• The system demonstrates robust performance in handling severe self-occlusion scenarios.

Conclusions:

• The developed probabilistic framework offers high-quality 3D human motion estimation, even with noisy depth data and self-occlusion.
• The local mixture of Gaussian Processes approach provides an efficient and adaptable solution for real-time applications.
• This advancement is beneficial for interactive applications like motion-based gaming and sports training.