SparseTrack: A Physics-Informed Transformer Framework for Real-Time Human Motion Reconstruction from Sparse IMUs.

Area of Science:

• Biomechanics
• Computer Vision
• Robotics

Background:

• Wearable inertial measurement units (IMUs) are crucial for human motion analysis.
• Current IMU systems often require numerous sensors, increasing cost and complexity.
• A need exists for sparse, efficient, and accurate IMU-based motion capture.

Purpose of the Study:

• To develop a sparse inertial human motion reconstruction framework using only five wearable sensors.
• To maintain real-time performance and biomechanical plausibility in motion capture.
• To enable practical applications like biomechanical digital twins.

Main Methods:

• Integration of Movella Xsens DOT IMUs with a learning-based inverse kinematics pipeline.
• Development of a real-time biomechanical digital twin for reconstruction and visualization.
• Training a sparse inference framework using the Virginia Tech Natural Motion Dataset and custom hard negative samples.

Main Results:

• Accurate full-body human motion reconstruction (excluding head) with a Mean Per-Joint Position Error of 5.96 cm using five sensors.
• Transformer-based temporal modeling outperformed recurrent and convolutional baselines in geometric accuracy and temporal smoothness.
• Physics-informed regularization and hard negative mining enhanced biomechanical consistency and reduced motion jitter.

Conclusions:

• The proposed framework enables accurate and real-time human motion reconstruction with a minimal sensor count.
• Transformer models and physics-informed techniques significantly improve motion capture quality.
• The system's interactive latency makes it suitable for biomechanical digital twin applications.