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

Wearable inertial measurement units are widely used for human motion analysis due to their portability; however, most IMU-based motion capture systems rely on dense sensor configurations that increase cost, complexity, and usability challenges in real-world applications. To address this limitation, this paper presents a sparse inertial human motion reconstruction framework that uses only five wearable sensors while maintaining real-time performance and biomechanical plausibility. The proposed framework integrates Movella Xsens DOT IMUs with a learning-based inverse kinematics pipeline and a real-time biomechanical digital twin for motion reconstruction and visualization. The evaluation was conducted in two phases: first, a real-time motion streaming system was established to validate sensor alignment, coordinate frame consistency, and end-to-end latency; second, a sparse inference framework was trained using the Virginia Tech Natural Motion Dataset combined with a custom dataset containing hard negative samples. Experimental results show that the system can accurately reconstruct full-body human motion, excluding head movement, with a local Mean Per-Joint Position Error of 5.96 cm using only five sensors. Comparative ablation studies demonstrate that Transformer-based temporal modeling achieves better geometric accuracy and temporal smoothness than recurrent and convolutional baselines, while physics-informed regularization and hard negative mining significantly improve biomechanical consistency and reduce motion jitter. Real-time experiments further demonstrate that the framework operates within interactive latency limits, highlighting its potential for biomechanical digital twin applications.