HARMES
收藏资源简介:
HARMES是由德国锡根大学和波恩大学联合推出的多模态可穿戴人类活动识别数据集,首次整合了腕部惯性测量单元(IMU)、大气环境传感器(湿度/温度/压力)和音频三种传感模态。该数据集包含20名参与者在家庭环境中执行的15类日常活动,总时长超过80小时,其中标记活动数据达61小时,每条记录包含141次独立活动重复。数据采集通过双腕佩戴智能设备实现,涵盖烹饪、清洁等复杂动作的异构传感信号。作为当前最大的腕部惯性-声学-环境传感联合数据集,其核心价值在于解决单一运动模态对相似活动(如洗手与刷牙)的识别模糊性问题,为医疗监护、智能家居等领域的多模态算法研究提供基准资源。
HARMES is a multimodal wearable human activity recognition dataset jointly launched by the University of Siegen and the University of Bonn in Germany. It is the first dataset to integrate three sensing modalities: wrist-worn inertial measurement units (IMUs), ambient environmental sensors (humidity, temperature and pressure), and audio. This dataset contains 15 categories of daily activities performed by 20 participants in a home environment, with a total duration of over 80 hours, among which labeled activity data totals 61 hours, and each record includes 141 independent activity repetitions. Data was collected via smart devices worn on both wrists, covering heterogeneous sensing signals from complex movements such as cooking and cleaning. As the largest currently available joint dataset of wrist-worn inertial, acoustic and environmental sensing, its core value lies in addressing the recognition ambiguity of similar activities (e.g., handwashing and toothbrushing) caused by single motion modality, providing benchmark resources for multimodal algorithm research in fields such as medical monitoring and smart homes.
数据集概述:HARMES
HARMES 是一个面向可穿戴人体活动识别(HAR)的多模态数据集。该数据集由20名受试者执行15种日常生活活动(ADL)时采集的数据组成,包含61小时的完整标注数据及额外1小时的自定义活动数据(总计20小时)。
数据版本
数据集提供两个版本:
- HARMES-RAW.zip:完整原始数据上传,包含所有未经重采样和缩放的原始传感器数据,数据已通过“ts_sync”时间戳同步。
- HARMES-preprocessed.pkl:Python pickle文件(每位受试者一个),包含分次记录的数据。数据已重采样至50Hz,并对IMU传感器值、BME280大气传感器值及音频数据进行了逐次Z-score归一化。
数据结构
原始数据结构
- 受试者文件夹(01-20)
- 记录文件夹(xy01-xy04,xy为受试者ID)
<timestamp1>_TASKLIST.csv:任务顺序清单(用于第1-3次记录)<timestamp2>.csv:活动标签及注释<rec_id>_merged.csv:来自Puck.js的数据(左手腕6轴IMU + 3轴BME280)recording_<date_str>.csv:智能手表IMU数据(右手腕6轴IMU)recording_<date_str>_SYNC_H5.h5:HDF5文件,包含右手腕麦克风录音数据(44.1kHz)
- 记录文件夹(xy01-xy04,xy为受试者ID)
预处理数据结构
- 文件:
participant_<受试者ID>.pkl - 内容:每位受试者的4-5次记录,每个记录为一个字典,包含以下键:
IMU_L(左手IMU数据,50Hz)IMU_R(右手IMU数据,50Hz)Audio(音频信号,44.1kHz)BME280(环境传感器数据,50Hz)Labels(标注的起始/结束时间,单位毫秒)rec_id(记录ID,与原始版本记录文件夹名一致)
关键词
- human activity recognition
- acoustic sensing
- IMU
- BME280
- Atmospheric Sensing
其他信息
- 许可协议:Creative Commons Attribution 4.0 International
- DOI:10.5281/zenodo.19425719
- 软件仓库:https://github.com/RBurchard/HARMES
- 开发语言:Python
- 开发状态:Active
数据文件
| 文件 | 大小 | MD5校验码 |
|---|---|---|
| HARMES-preprocessed.zip | 22.0 GB | eda78bd7c5915a3cbe438fe05a1f0b3b |
| HARMES-RAW.zip | 19.9 GB | b4c041f3ab1f139d554a9d2615639682 |

- 1HARMES: A Multi-Modal Dataset for Wearable Human Activity Recognition with Motion, Environmental Sensing and Sound锡根大学; 波恩大学·Lamarr机器学习和人工智能研究所 · 2026年



