Three-Dimensional Human Pose Estimation from Sparse IMUs through Temporal Encoder and Regression Decoder.

Area of Science:

• Computer Vision
• Human Motion Analysis
• Biomechanical Engineering

Background:

• Three-dimensional (3D) pose estimation is crucial for human motion analysis, with inertia-based methods gaining traction.
• Current inertial measurement unit (IMU)-based systems often involve dense sensors and complex calibration, limiting natural movement.
• Sparse IMU methods show promise but struggle with accuracy issues like body shaking and movement ambiguity.

Purpose of the Study:

• To enhance 3D human pose estimation accuracy using sparse IMUs.
• To address limitations of existing temporal-feature-based methods in sparse IMU pose estimation.
• To develop a fusion approach combining temporal and spatial features for robust pose prediction.

Main Methods:

• A multistage encoder-decoder network was designed, incorporating a temporal convolutional encoder and a human kinematics regression decoder.
• The approach fuses temporal feature information with human kinematic feature information for pose prediction.
• Experiments were conducted on benchmark datasets to validate the proposed method.

Main Results:

• The proposed method achieved a 13.6% reduction in mean per joint position error on the Total Capture dataset.
• A 19.4% decrease in mean per joint position error was observed on the DIP-IMU dataset.
• Quantitative comparisons confirmed the superiority of the fused temporal and kinematic features over existing methods.

Conclusions:

• The fusion of temporal and spatial features significantly improves 3D human pose estimation accuracy with sparse IMUs.
• The developed encoder-decoder network effectively leverages human kinematic topology for enhanced pose prediction.
• This approach offers a more accurate and less intrusive solution for 3D human motion analysis.