Rotation-Invariant Multi-IMU Activity Recognition under Independent Per-Location Orientation Shifts
TRI-HAR is a rotation-invariant framework for multi-IMU Human Activity Recognition that handles independent orientation offsets between different body locations structurally. It reshapes sensor data into triaxial vectors, uses a shared SO(3)-equivariant backbone, and fuses location-level invariant features to achieve robustness without needing rotation augmentation or calibration.