josefheidler/har_older-adults_2023-harth

HAR70+ — Older Adult Activity Recognition (Thigh + Back Accelerometry)

Two Axivity AX3 accelerometers (right thigh and lower back, 50 Hz) worn during semi-structured free-living sessions by 18 older adults (age 70–95) in Norway. Harmonized from the NTNU HAR70+ release into per-subject parquet — one row per raw sample. 18 participants, 2.26 M samples, 100% labelled, 12.6 h labeled.

Source

• Paper: Ustad, A., Logacjov, A., Trollebø, S. Ø., Thingstad, P., Vereijken, B., Bach, K., & Mork, P. J. (2023). Validation of an Activity Type Recognition Model Classifying Daily Physical Behavior in Older Adults. Sensors, 23(1), 130. https://doi.org/10.3390/s23010130
• Raw data: https://github.com/ntnu-ai-lab/harth-ml-experiments/tree/main/har70plus
• License: CC-BY-4.0

Protocol

• Participants: 18 older adults, age 70–95 (five used walking aids), recruited from a rehabilitation clinic in Trondheim, Norway.
• Sensor: Axivity AX3 triaxial accelerometer (±8 g, 50 Hz), placed on the right thigh (≈10 cm above the kneecap) and lower back (≈3rd lumbar vertebra), USB connector pointing down.
• Protocol: ≈40-minute semi-structured free-living session per participant. Participants performed their usual daily activities while a trained observer accompanied them and annotated behaviour in real time. Seven activity types were annotated: walking, shuffling, stair climbing (ascending and descending), standing, sitting, and lying.

Ground-truth provenance

Activities were annotated by a trained observer accompanying each participant during the session using a mobile logging application. Timestamps were recorded manually. No video annotation was used; inter-rater reliability is not reported in the paper. All 2,259,597 samples are labeled — there are no unlabeled transition windows.

Schema

All timestamps are tz-aware Europe/Oslo.

column	dtype	notes
subject	string	participant alias (e.g. agile-anteater)
timestamp	datetime64[ns, Europe/Oslo]	real wall-clock time of the recording, tz-aware Europe/Oslo
thigh_acc_x	float64	right-thigh acceleration, corrected frame (see Axis orientation), units = g
thigh_acc_y	float64	right-thigh acceleration, corrected frame, units = g
thigh_acc_z	float64	right-thigh acceleration, corrected frame, units = g
back_acc_x	float64	lower-back acceleration, corrected frame (see Axis orientation), units = g
back_acc_y	float64	lower-back acceleration, corrected frame, units = g
back_acc_z	float64	lower-back acceleration, corrected frame, units = g
label	string (nullable)	activity base (walk, shuffle, stairs, stand, sit, lie)
variant	string (nullable)	direction modifier (ascending, descending); null for most activities

Example rows

subject          timestamp                       thigh_acc_x  thigh_acc_y  thigh_acc_z  back_acc_x  back_acc_y  back_acc_z  label   variant
agile-anteater   2021-03-24 14:42:03.839+01:00   0.980       -0.112        0.048         0.999       0.063       0.141      stand   <NA>
agile-anteater   2021-03-24 14:43:32.599+01:00   0.987       -0.219       -0.000         1.042       0.083      -0.181      walk    <NA>
agile-anteater   2021-03-24 14:53:31.280+01:00   0.982       -0.103        0.196         0.989      -0.094      -0.261      stairs  descending

Label vocabulary

label and variant are stored in separate columns. variant is null for all activities except stairs.

label	variant	meaning	labeled min
lie	—	lying	67.7
shuffle	—	shuffling / slow side-step	22.0
sit	—	sitting	161.2
stairs	ascending	stair climbing upward	1.5
stairs	descending	stair climbing downward	1.7
stand	—	standing	139.4
walk	—	walking	359.8

Total labeled: 753.3 min / 12.6 h across 18 participants.

Coverage

All 2,259,597 rows are labeled — this dataset has no unlabeled transition windows. Filter to labeled rows (no-op but safe to use consistently across the hub):

df_labeled = df[df["label"].notna()]

Axis orientation

All accelerometer columns use the hub standard axis convention, shared across all datasets on this hub:

• x runs along the body segment toward the head — reads +1 g when the person stands upright, drops toward 0 when they lie down.
• y points to the person's right — positive when tilting right, negative when tilting left.
• z points forward — positive when leaning or stepping forward, negative when leaning backward.
• At rest standing upright: x ≈ +1 g, y ≈ 0, z ≈ 0.

Thigh sensor (right thigh, ≈10 cm above the kneecap, USB connector pointing down):

Native frame (standing): x=down, y=right, z=backward. Correction: negate x and z (keep y). Hub frame: x=up, y=right, z=forward.

Back sensor (lower back, ≈3rd lumbar vertebra, USB connector pointing down):

Native frame (standing): x=down, y=left, z=forward. Correction: negate x and y (keep z). Hub frame: x=up, y=right, z=forward.

The raw native-frame values are available in the source CSV files at the GitHub link above.

Harmonization notes

No participants were excluded. All 18 participants from the GitHub release are included.

Use

Intended for human activity recognition (HAR) benchmarking on older-adult populations. Unlike the HARTH adult dataset, HAR70+ contains only free-living sessions with no controlled laboratory stages, making it suitable for evaluating model generalization to natural movement in this cohort — including shuffling gait patterns typical of older adults. Evaluate with leave-one-subject-out (LOSO) cross-validation; avoid random row-level splits which would leak temporal context across subjects.

Loading

from datasets import load_dataset

ds = load_dataset("josefheidler/har_older-adults_2023-harth")
df = ds["train"].to_pandas()

Load a single subject:

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/josefheidler/har_older-adults_2023-harth/harmonized/agile-anteater.parquet"
)

Citation

Ustad, A., Logacjov, A., Trollebø, S. Ø., Thingstad, P., Vereijken, B., Bach, K., & Mork, P. J. (2023). Validation of an Activity Type Recognition Model Classifying Daily Physical Behavior in Older Adults. Sensors, 23(1), 130. https://doi.org/10.3390/s23010130