Human motion tracking by temporal-spatial local gaussian process experts.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Human pose estimation from motion tracking is a challenging regression problem.
• Modeling the high-dimensional, multimodal conditional distribution from observation to state space is complex.
• Existing methods struggle with multimodality due to large training sample requirements.

Purpose of the Study:

• To propose a novel online sparse Gaussian Process (GP) regression model for 3-D human motion recovery in monocular videos.
• To address the limitations of existing techniques in handling multimodal distributions.
• To develop a robust and efficient system for nonlinear human motion tracking.

Main Methods:

• Developed an online sparse Gaussian Process (GP) regression model.
• Implemented a local mixture of GP experts, combining temporal and spatial information.
• Created a hybrid system with self-initialization for seamless integration.
• Defined local experts online within small neighborhoods for efficient learning and inference.

Main Results:

• The proposed model effectively recovers 3-D human motion from monocular videos.
• The system demonstrates robustness in tracking nonlinear human movements.
• Achieved significant performance improvements compared to existing models on real-world datasets (HumanEva, PEAR).

Conclusions:

• The novel online sparse GP regression model offers an effective solution for 3-D human motion estimation.
• The hybrid system combining temporal and spatial experts successfully addresses multimodality.
• The method provides efficient learning and inference, outperforming prior approaches in visual tracking accuracy.