josefheidler/har_adults_2021-harth

HARTH — thigh + back accelerometry, healthy adults, activity recognition

Dual-sensor accelerometry (Axivity AX3, 50 Hz, ±8 g) from the right thigh and lower back in 31 healthy adults, with activity ground truth for 12 activity types including walking, running, stair climbing, cycling, and postural activities. Recordings are 1–2.5 hours per participant of semi-free-living activity. Harmonized from the HARTH GitHub release into per-subject parquet — one row per raw accelerometer sample (50 Hz). 31 participants, 8.3 M samples, >99.9 % labelled, 46.2 h labeled.

Designed to resemble the HUNT4 accelerometer data (Trøndelag Health Study, Norway, ≈35,000 participants using the same dual-sensor setup at thigh and lower back), it serves as a labelled training resource for large-scale population health research.

Source

• Paper: Logacjov, A., Bach, K., Kongsvold, A., Bårdstu, H. B., & Mork, P. J. (2021). HARTH: A Human Activity Recognition Dataset for Machine Learning. Sensors, 21(23), 7853. https://doi.org/10.3390/s21237853
• Related paper: Bach, K., Kongsvold, A., Bårdstu, H., Bardal, E. M., Kjærnli, H. S., Herland, S., Logacjov, A., & Mork, P. J. (2022). A Machine Learning Classifier for Detection of Physical Activity Types and Postures During Free-Living. Journal for the Measurement of Physical Behaviour, 5(1), 24–31. https://doi.org/10.1123/jmpb.2021-0015
• Raw data: https://github.com/ntnu-ai-lab/harth-ml-experiments/tree/main/harth (original UCI release: https://archive.ics.uci.edu/dataset/779/harth — 22 participants; the GitHub version used here includes 31 participants with corrected labels)
• License: MIT

Protocol

• Participants: 31 healthy adults (8 female, 14 male in original 22-subject release; age 38.6 ± 14 y, height 177.3 ± 8.3 cm, weight 72.9 ± 10.6 kg, BMI 23.1 ± 2.3) (range: age 25–68 y, weight 56–92 kg, height 157–191 cm, BMI 19.2–28.4 kg/m²).
• Sensors: Two Axivity AX3 per participant — right thigh (≈10 cm above the upper border of the patella) and lower back (3rd lumbar vertebra), both mounted with USB connector pointing down. Attachment: a 5 × 7 cm moisture-permeable film (Opsite Flexifix, Smith & Nephew) on the skin, sensor fixed with double-sided tape, covered by a second 10 × 8 cm film layer.
• Sample rate: 50 Hz (100 Hz raw, published at 50 Hz).
• Recording — two sessions:
  • Session 1 (22 participants, 6 female reported in protocol description): 1.5–2 h at the participant's workplace or home (semi-free-living). Participants were asked to accumulate at least 2–3 min each of sitting, standing, lying, walking, and running/jogging; no enforced order. Total: ≈1,804 min (≈30 h) across 22 participants (average 120 ± 21.6 min/participant).
  • Session 2 (9 additional participants): 60 ± 9 min per participant (417.6 min / ≈7 h total), focused on walking, running, and cycling across flat, uphill, and downhill sections; incidental postures and stair-climbing also annotated.
  • Raw data sampled at 100 Hz; downsampled to 50 Hz for the public release.

Ground-truth provenance

Frame-by-frame video annotation using ANVIL software. Ground truth recorded with a GoPro Hero 3+ chest-mounted camera (30 fps, 640 × 360 pixels) pointing toward the feet. Two independent human experts annotated each file; Fleiss' κ = 0.96. Sensor and video signals were aligned using three heel drops visible in both streams.

Published baseline (Logacjov et al. 2021, 12 labels, leave-one-subject-out): SVM F1 = 0.66 (best); after merging to 9 activity groups: SVM F1 = 0.81. Bach et al. 2022 (6 merged labels, dual-sensor XGBoost, 5-s window): overall accuracy 96 %, κ = 0.92.

Schema

timestamp is epoch-anchored elapsed time: every recording is normalised so its first sample falls at 1970-01-01 00:00:00 UTC. The date is a sentinel — the Axivity AX3 RTC was not synchronised to wall-clock for all recordings. Within-participant relative timing is accurate at 50 Hz.

column	dtype	notes
subject	string	participant alias (e.g. amber-antelope)
timestamp	datetime64[ns, UTC]	epoch-anchored elapsed time (see note above)
thigh_acc_x	float64	right-thigh acceleration, corrected frame, g
thigh_acc_y	float64	right-thigh acceleration, corrected frame, g
thigh_acc_z	float64	right-thigh acceleration, corrected frame, g
back_acc_x	float64	lower-back acceleration, corrected frame, g
back_acc_y	float64	lower-back acceleration, corrected frame, g
back_acc_z	float64	lower-back acceleration, corrected frame, g
label	string (nullable)	activity base (e.g. walk, bicycle)
variant	string (nullable)	posture/direction modifier; null when unspecified

Example rows:

subject           timestamp                thigh_acc_x  thigh_acc_y  thigh_acc_z  back_acc_x  back_acc_y  back_acc_z  label    variant
amber-antelope    1970-01-01 00:00:12      0.95         -0.11        0.07         0.92        0.04        -0.02       stand    null
amber-antelope    1970-01-01 00:05:30      0.62         -0.08        0.43         0.61        0.03        0.22        walk     null
amber-antelope    1970-01-01 00:22:10      0.58         -0.06        0.51         0.55        0.02        0.31        run      null
amber-antelope    1970-01-01 00:35:00      0.97         -0.03        0.04         0.94        0.01        0.02        lie      null
amber-antelope    1970-01-01 00:50:20      0.73         -0.04        0.39         0.70        0.02        0.18        bicycle  pedalling-seated
wry-wolverine     1970-01-01 00:06:02      0.84         -0.09        0.12         0.81        0.02        0.09        null     null

Note: acc values above are illustrative — read actual parquet for true values.

Label vocabulary

label and variant are stored in separate columns. variant is null for activities with no posture or direction modifier.

label	variant	meaning	labeled min
bicycle	coasting-seated	Freewheeling (not pedalling), seated	13.9
bicycle	coasting-standing	Freewheeling (not pedalling), standing	2.6
bicycle	pedalling-seated	Active pedalling, seated on saddle	170.4
bicycle	pedalling-standing	Active pedalling, standing out of saddle	18.7
lie	—	Lying on stomach, back, or side	196.0
run	—	Both feet leave the ground each stride	115.2
shuffle	—	Stepping in place / turning on the spot	105.6
sit	—	Buttocks on a seat, bed, or floor	1,097.4
stairs	ascending	Stair climbing upward	31.7
stairs	descending	Stair climbing downward	29.3
stand	—	Upright, stationary, feet supporting weight	420.8
walk	—	Locomotion, one stride or more, any direction	572.3

Total labeled: 2,774 min / 46.2 h across 31 participants.

Coverage

Nearly all samples are labeled (99.99 %, 8,321,874 of 8,322,728 rows). The only unlabeled rows (854 rows, ≈17 s) appear in wry-wolverine where label code 17 (undocumented in the upstream release) is mapped to null. Filter with df[df['label'].notna()] to keep only labeled rows, or work with the full dataset directly — the effect is negligible.

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 upper border of the patella, USB connector pointing down):

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

Back sensor (lower back, L3, USB connector pointing down):

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

The native axes are confirmed in Logacjov et al. 2021: "seen from the participant's perspective while standing upright, the lower back sensor's x-axis points downward, the y-axis to the left, and the z-axis forward. For the thigh sensor, the y-axis points to the right and the z-axis backward."

Harmonization notes

Label codes 10 (transport sitting) and 11 (transport standing) — present in 4 subjects, ≈2.4 min combined — are collapsed to plain sit and stand respectively. Label code 17 (undocumented, ≈17 s in wry-wolverine only) is mapped to null. No participants excluded.

Use

Intended for human activity recognition (HAR) from dual-sensor thigh + back accelerometry — predicting label (12 classes) from thigh_acc_x/y/z and back_acc_x/y/z. The standard evaluation protocol is leave-one-subject-out (LOSO) cross-validation across 31 participants. Avoid random row-level splits — they leak temporal and subject-level context across folds.

Loading

Load all subjects into one table:

from datasets import load_dataset

ds = load_dataset("josefheidler/har_adults_2021-harth")
df = ds["train"].to_pandas()

Load a single subject:

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/josefheidler/har_adults_2021-harth/harmonized/amber-antelope.parquet"
)

Citation

Logacjov, A., Bach, K., Kongsvold, A., Bårdstu, H. B., & Mork, P. J. (2021). HARTH: A Human Activity Recognition Dataset for Machine Learning. Sensors, 21(23), 7853. https://doi.org/10.3390/s21237853