TRI-HAR: A Rotation-Invariant Framework for Multi-IMU Human Activity Recognition

A recent study published on arXiv (2608.15621) presents TRI-HAR (Truly Rotation-Invariant HAR), a framework aimed at enhancing the robustness of human activity recognition (HAR) against independent orientation changes of inertial measurement units (IMUs) positioned on various body parts. This challenge is particularly evident in self-administered devices, like those used for home rehabilitation and exercise tracking, where users frequently reposition IMUs, leading to orientation discrepancies that traditional scalar HAR models struggle with. Current methods often depend on rotation augmentation or calibration processes, which require specific reference frames. In contrast, TRI-HAR integrates rotation invariance into the model's structure by transforming accelerometer and gyroscope data into triaxial vectors, employing a shared SO(3)-equivariant backbone, and applying invariant projections at each IMU site. This innovation negates the need for calibration or augmentation, providing a more dependable solution for real-world applications. The implications of this research are significant for the expanding domains of wearable technology and digital health, where precise activity recognition is vital for tracking patient progress and ensuring proper exercise execution. Additionally, the framework may be applicable to other sensor-based recognition challenges beyond HAR.