ElderHAR: A Real-World Inertial Dataset for Elderly Activity Recognition

Background and Purpose:

Falls are a major health concern among older adults, often leading to disability, hospitalization, and loss of independence. Human Activity Recognition (HAR) using wearable sensors has emerged as a promising tool for continuous fall risk assessment, but existing public datasets largely rely on young, healthy participants and controlled laboratory tasks. This gap limits the development of models that generalize well to elderly populations and real-world conditions.

To address this need, we present ElderHAR, a dataset of continuous, real-world inertial recordings from 54 older adults performing daily activities and natural transitions across six care institutions. Collected with a single waist-mounted IMU, the dataset comprises over 2.6 million labeled samples, rich metadata, and supporting code for preprocessing and benchmarking. By emphasizing ecological validity and elderly-specific variability, ElderHAR provides a valuable resource for advancing HAR, fall risk assessment, and elderly care technologies.

Dataset Content:
The dataset comprises approximately 15 hours of annotated human activity data, amounting to over 2.6 million samples of synchronized tri-axial accelerometer and gyroscope signals. Data were collected at a fixed sampling rate of 50 Hz and were stored in comma-separated values (CSV) format files. Each file corresponds to a single trial (a single continuous activity circuit) performed by a participant.

Each CSV file contains the following columns:

• Column 1: Sample — sample count; 
• Column 2: Subject — subject ID; 
• Column 3: Trial — trial ID; 
• Column 4-6: AccX, AccY, and AccZ — acceleration in the X, Y, and Z axes (in g); 
• Column 7-9: GyrX, GyrY, and GyrZ — angular velocity in the X, Y, and Z axes (in °/s); and 
• Column 10: Label — numeric code corresponding to the activity class.

A separate metadata CSV file is also provided, containing relevant information for each participant, including, when available: (i) subject ID, (ii) age, (iii) gender, (iv) height (cm), (v) weight (kg), (vi) abdominal circumference (cm), and (vii) faller status, indicated as Faller (F) or Non-Faller (NF).

Additional documentation is provided in an accompanying README.txt file located in the main dataset folder. This file includes essential information on the data acquisition protocol, sensor specifications, label ID-to-activity mapping, and subject metadata. It also offers guidance on data usage, including preprocessing steps and modeling considerations, to support researchers in effectively working with the dataset.

Files follow the naming convention SubjectID_TrialNumber.csv. For example, S12_T01.csv corresponds to the first trial of participant 12. Each subject performed only one trial.

Participants & Data Acquisition System:
Fifty-four participants from six elderly care institutions agreed to participate in the data acquisition. The recruitment aimed to maximize the diversity in physical stature, mobility characteristics, and social backgrounds. To achieve this, participants were recruited from both public and private institutions, including individuals with different educational backgrounds, and efforts were made to balance gender representation, although a higher proportion of women was ultimately included. The inclusion criteria were: (i) community-dwelling and able to walk independently (FAC > 3), (ii) no severe cognitive impairment (MMSE > 15), and (iii) able to follow verbal instructions.

Clinicians and caregivers assessed potential participants using clinical scales to ensure eligibility. Demographic and physiological data, including age, height, weight, gender, abdominal circumference, and faller status, were recorded. For a subset of participants (n = 32) from three institutions, abdominal circumference was available, ranging from 83 cm to 125 cm, further illustrating the variability in body composition across the sample. Table 1 summarizes the main characteristics of the study population.

Data were collected using an instrumented waistband, originally developed by our team as described by Gonçalves et al. (2018). This system included an LSM6DSOX inertial sensor (accelerometer ±16 g, gyroscope ±2000 º/s) integrated into a compact, lightweight module. The sensing unit was centrally located on the waistband, connected to a control unit powered by an STM32F4 Discovery board, which managed data collection at 50 Hz for all sessions. Data were stored on a USB drive and an HC06 Bluetooth module enabled communication with external devices.

The IMU was strategically placed on the lower back via a waistband to optimally capture trunk movements, essential for postural transitions, gait and other daily activities. This placement ensures high-quality data acquisition while maintaining participant comfort and minimizing interference with natural movements.

To facilitate data acquisition and labeling, two custom Android applications were developed: one for managing the acquisition process and another for annotate each activity start timestamp. The waistband fabric was selected to improve comfort and adjustability for diverse abdominal circumferences, ensuring suitability for all elderly participants during the data acquisition sessions.

Table 1 - Participants' demographic information

Data Collection Methodology:

The experimental protocol was designed to integrate four essential components: (i) precise control over stimuli, (ii) high reproducibility of the experimental conditions, (iii) preservation of ecological validity, and (iv) promotion of real-world learning transfer.

Data were collected in a continuous circuit format, with real-world settings and without movement restrictions, ensuring that the captured behaviors were natural and reflective of typical daily living. The dataset includes nine key activities relevant for FRA: a) Walking, b) Standing, c) Sitting in a chair, d) Laying in bed, e) Walking upstairs, f) Walking downstairs, g) Picking objects from the ground, h) Reaching for objects at height (to standardize movement, all participants lifted their heels off the ground while fully extending one or both arms upward), and i) Turning while walking or standing. Additionally, four important transitions, namely sit-to-stand, lay-to-sit, and their respective reverses, were also performed by the participants during the proposed circuit, making a total of 13 activities recorded.

The protocol included: (i) a pre-defined circuit of activities of daily living (ADLs); (ii) a minimum duration of two minutes for each cyclical activity (e.g., walking) and static posture (e.g., standing); (iii) naturally shorter, non-time-bound transitions (e.g., sit-to-stand); (iv) activity start and end times manually labeled in real time by a test administrator using a synchronized smartphone app; and (v) flexibility to pause the circuit if participants showed fatigue, as well as to skip activities they were uncomfortable performing.

Throughout the circuit, the administrator provided clear verbal instructions to guide participants through each activity and transition — for example, “walk to the chair and sit down normally” or “pick up the object from the floor and stand back up” — as well as directional cues such as “turn right” or “walk towards the stairs” to adapt to each institution’s layout.
Activity sequences and durations varied across institutions based on available space and layout, adding diversity to the dataset. 

Data Labeling and Synchronization

During data acquisition, a trained administrator used a custom smartphone application to manually mark the start of each activity in real-time. This app was synchronized with the inertial sensing unit via timestamp alignment, ensuring that all activity annotations were accurately time-stamped to the millisecond. Each of the 13 predefined activities — ranging from static postures to cyclical movements and transitions — was labeled in this manner as it was performed.

Following the acquisition sessions, the recorded sensor data were post-processed and annotated on a per-sample basis. Using the synchronized timestamps between the sensor logs and manual annotations, each sample in the dataset was assigned one of the 13 activity labels listed in Table 2. These include static postures (standing, sitting, lying), dynamic movements (walking, stair ascent/descent), object-related tasks (picking and reaching), directional changes (turning), and transitions (sit-to-stand, stand-to-sit, lying-to-sit, and sit-to-lying).

After the initial annotations, all labels were also manually adjusted to remove any potential deviations or errors introduced during data collection. This manual optimization was particularly important for transitional activities, such as sit-to-stand or stand-to-sit movements. The start and end times of these transitions were normalized across all participants to ensure consistency in the dataset. Given the absence of video recordings due to privacy reasons, the optimization was based entirely on sensor data. Although this approach presented challenges, it ensured that participants' privacy was maintained while still capturing high-quality, labeled activity data.

Table 2 - Labels and corresponding activities in the dataset