Fast Human Motion reconstruction from sparse inertial measurement units considering the human shape.

Area of Science:

• Computer Vision
• Biomechanical Engineering
• Machine Learning

Background:

• Inertial Measurement Unit (IMU)-based motion capture offers potential for complex environments.
• Sparse IMU methods are flexible but face challenges in computational efficiency and latency.
• Accurate and efficient full-body motion estimation is crucial for various applications.

Purpose of the Study:

• To propose Fast Inertial Poser, a novel deep neural network for efficient full-body motion estimation using sparse IMUs.
• To address the computational efficiency and latency limitations of existing sparse IMU-based methods.
• To improve the accuracy and real-time performance of human motion reconstruction.

Main Methods:

• Developed a deep neural network leveraging recurrent neural networks and a kinematics tree structure.
• Incorporated human body shape information and utilized causal observations to eliminate future frame dependency.
• Employed separate network modules for upper and lower body joint estimation.
• Implemented a single-frame kinematics inverse solver for joint rotation estimation.

Main Results:

• Fast Inertial Poser significantly enhances inference speed and reduces latency compared to prior methods.
• The model achieves high reconstruction accuracy while maintaining efficiency.
• Demonstrated real-time performance with 65 frames per second (fps) and 15 milliseconds (ms) latency on an embedded computer.

Conclusions:

• Fast Inertial Poser provides an efficient and accurate solution for sparse IMU-based full-body motion estimation.
• The proposed method overcomes key challenges in computational efficiency and latency.
• This advancement holds promise for real-time motion capture in diverse applications.