跳转至

文章背景与核心概要

本文介绍了 TRI-HAR（真正的旋转不变性人类活动识别） 框架，旨在解决使用多传感器可穿戴设备（IMU）进行人类活动识别（HAR）时面临的旋转偏移难题。在居家康复或运动监测等应用场景中，用户在不同会话间重新佩戴可穿戴设备往往会导致身体各个部位产生独立的朝向偏移。传统的标量 HAR 模型在没有大量旋转数据增强、显式校准流程或严格参考系假设的情况下，很难有效应对这一问题。

为了克服这些局限性，TRI-HAR 通过将传感器数据转换为三轴向量，直接在模型架构中构建了旋转不变性。该模型通过共享的 \(\text{SO}(3)\)-等变骨干网络处理这些向量，并在每个位置应用不变投影，最后融合特征进行分类。在四个多IMU基准数据集上的实验表明，TRI-HAR 在固定的各位置独立旋转下能够成功保持 macro-F1 性能，并且在无需显式数据增强的情况下，表现优于依赖旋转增强的基线模型。

Rotation-Invariant Multi-IMU Activity Recognition under Independent Per-Location Orientation Shifts

Summary

This paper presents TRI-HAR (Truly Rotation-Invariant HAR), a novel framework for Human Activity Recognition (HAR) using multi-sensor wearable devices (IMUs). Reattaching wearables across sessions—such as in at-home rehabilitation or exercise monitoring—often introduces independent orientation shifts at different body locations. Traditional scalar HAR models struggle with this without extensive rotation augmentation, explicit calibration procedures, or rigid reference-frame assumptions. TRI-HAR builds rotation invariance directly into the model's architecture by transforming sensor data into triaxial vectors, processing them through a shared \(\text{SO}(3)\)-equivariant backbone with invariant projection per location, and fusing the features for classification. Experiments across four multi-IMU benchmarks demonstrate that TRI-HAR successfully preserves macro-F1 performance under fixed independent per-location rotations and outperforms rotation-augmented baselines without requiring explicit data augmentation.

Metadata

  • arXiv ID: arXiv:2608.15621 [cs.AI]
  • Related DOI: 10.1145/3830727.3834824
  • Conference: ACM International Symposium on Wearable Computing (ISWC) 2026
  • Subjects: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
  • Submission Date: August 16, 2026

Authors

  • Seungyeol Baek
  • Yoonbyung Chai
  • Yonghyeon Lee
  • Sungjoon Choi
  • Sungho Suh

Abstract

使用自主佩戴的可穿戴设备（如居家康复和运动监测）进行人类活动识别（HAR），通常需要在不同会话之间重新佩戴惯性测量单元（IMU）。在多IMU设置中，这可能会导致身体不同部位产生独立的方向偏移，这种部署偏移是常规标量 HAR 模型无法在结构上处理的。

Human Activity Recognition (HAR) with self-administered wearables, such as at-home rehabilitation and exercise monitoring, often requires reattaching inertial measurement units (IMUs) across sessions. In multi-IMU settings, this can induce independent orientation offsets across body locations, a deployment shift that conventional scalar HAR models do not structurally handle.

现有的解决方法依赖于旋转数据增强（其鲁棒性取决于采样的变换），或者依赖需要额外参考系假设或显式流程的校准与方向归一化流水线。我们提出了 Truly Rotation-Invariant HAR (TRI-HAR)，这是一个旋转不变的框架，它将对独立各位置 IMU 方向偏移的鲁棒性变成了一种模型的结构属性。TRI-HAR 将加速度计和陀螺仪的数据流重塑为三轴向量，对每个 IMU 位置应用共享的 \(\text{SO}(3)\)-等变骨干网络和不变投影，并融合所产生的不变特征进行活动分类。在四个多IMU基准测试中，TRI-HAR 在固定的各位置独立 \(\text{SO}(3)\) 旋转下保持了 macro-F1 性能，并且在无需旋转数据增强的情况下，在此类目标偏移下优于旋转增强的基线模型。

Existing remedies rely on rotation augmentation, whose robustness depends on sampled transformations, or calibration and orientation-normalization pipelines requiring additional reference-frame assumptions or explicit procedures. We present Truly Rotation-Invariant HAR (TRI-HAR), a rotation-invariant framework that makes robustness to independent per-location IMU orientation offsets a structural model property. TRI-HAR reshapes accelerometer and gyroscope streams into triaxial vectors, applies a shared \(\text{SO}(3)\)-equivariant backbone and invariant projection to each IMU location, and fuses the resulting invariant features for activity classification. Across four multi-IMU benchmarks, TRI-HAR preserves macro-F1 under fixed independent per-location \(\text{SO}(3)\) rotations and outperforms rotation-augmented baselines under this target shift without requiring rotational augmentation.

Access & Resources