josefheidler/har_children_2024-harth

Tørring 2024 — thigh + back accelerometry, children (typically developing + cerebral palsy), activity recognition

Dual-sensor accelerometry (Axivity AX3, 50 Hz, ±8 g) from the thigh and lower back in children with and without cerebral palsy, with activity ground truth for 13 activity types including walking, running, jumping, cycling, and postural activities. Recordings were collected across lab, gymnasium, and outdoor settings. Harmonized from the Dataverse release into per-subject parquet — one row per raw accelerometer sample (50 Hz). 78 participants (46 TD, 32 CP), 11.7 M samples, 100% labelled, 65.0 h labeled.

Source

• —Paper: Tørring, B., Refaat, M., Helbostad, J. L., Adde, L., & Ustad, A. (2024). Validation of two novel human activity recognition models for typically developing children and children with cerebral palsy. Gait & Posture, 108, 131–137. https://doi.org/10.1016/j.gaitpost.2023.11.013
• —Raw data: https://doi.org/10.18710/EPCXCC (DataverseNO)
• —License: CC0 1.0

Protocol

• —Participants (TD): 63 typically developing children in the study (46 in this release — see Harmonization notes). 35 male / 28 female, age 10.5 ± 2.6 years (range 6–15).
• —Participants (CP): 16 in the original paper (32 in this release — see Harmonization notes). 8 male / 8 female, age 11.4 ± 2.2 years (range 8–17), GMFCS levels I–II.
• —Sensors: Two Axivity AX3 accelerometers, ±8 g, 50 Hz per participant — anterior midline thigh (midpoint between ASIS and proximal patella; least-affected side for CP participants) and lower back (~L3 vertebra), both mounted with USB connector pointing down.
• —Sample rate: 50 Hz (TD recorded at 200 Hz, CP at 100 Hz; the Dataverse release is already downsampled to 50 Hz — no further resampling applied).
• —Protocol: Structured sessions conducted in a laboratory, gymnasium, and outdoors. Activities included individual and group sessions with both standardised tasks (walking, running, stair climbing, cycling) and free-play activities (jumping, shuffling, play). Each session covered a predefined set of activities; duration per activity varied by participant.

Ground-truth provenance

Video annotation using GoPro Hero 3+ cameras mounted in room corners, with frame-by-frame coding in Anvil v6. Four independent raters annotated the recordings; inter-rater reliability exceeded 0.95. Sensor–video synchronisation was achieved by the researcher flicking the accelerometer three times at the start of each session, producing a distinct acceleration peak visible in both streams.

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 reliably synchronised to wall-clock. Within-participant relative timing is accurate at 50 Hz.

Example rows:

subject              timestamp                thigh_acc_x  thigh_acc_y  thigh_acc_z  back_acc_x  back_acc_y  back_acc_z  label    variant   cohort
agile-axolotl        1970-01-01 00:02:10      0.97         -0.02        0.03         0.97        0.01        -0.01       stand    null      td
agile-axolotl        1970-01-01 00:08:45      0.63         -0.07        0.45         0.61        0.03        0.22        walk     null      td
young-reedbuck       1970-01-01 00:15:20      0.71         -0.05        0.38         0.68        0.02        0.19        bicycle  seated    cp
bold-jacamar         1970-01-01 00:31:00      0.97         -0.01        0.02         0.97        0.00        0.01        stand    null      td
young-reedbuck       1970-01-01 00:45:30      null         null         null         null        null        null        null     null      cp

Note: acc values above are illustrative — read the 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.

Total labeled: 3,902 min / 65.0 h across 78 participants.

Coverage

All 11,706,932 rows (100%) carry a label — there are no unlabeled rows in this dataset. Filter with df[df['label'].notna()] if your pipeline expects an explicit filter; it will return the full dataset unchanged.

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 (anterior midline thigh, 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.

Because the thigh is mounted anterior-midline (not on the right thigh as in the HARTH adults), its native frame matches this dataset's back sensor, and the correction is the same (negate x and y, keep z) — not the HARTH-adults thigh recipe (negate x, z). In addition, a per-subject orientation normalization enforces the hub convention against genuine mounting variation baked into the raw data: subjects mounted with the thigh rotated 180° about its long axis (mostly CP participants, sensor on the least-affected leg — anterior would otherwise read posterior) are rotated back, and one subject mounted 180° along the leg is corrected to long-axis-up. Verified: every subject reads sitting thigh-z positive and standing thigh-x positive, matching the other datasets.

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.

For the native raw values, refer to the Dataverse release.

Harmonization notes

Participant count: The original paper reports 16 CP and 63 TD participants. This release contains 32 CP and 46 TD (78 total) — the Dataverse release includes 32 numeric (CP) files rather than 16, and 22 TD files are absent from the release and excluded here.

Use

Intended for human activity recognition (HAR) from dual-sensor thigh + back accelerometry in children — predicting label (13 classes) from thigh_acc_x/y/z and back_acc_x/y/z. The cohort column enables separate evaluation for typically developing children and children with cerebral palsy. The recommended evaluation protocol is leave-one-subject-out (LOSO) cross-validation, applied within or across cohorts depending on the research question. 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_children_2024-harth")
df = ds["train"].to_pandas()

Load a single subject:

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/josefheidler/har_children_2024-harth/harmonized/agile-axolotl.parquet"
)

Citation

Tørring, B., Refaat, M., Helbostad, J. L., Adde, L., & Ustad, A. (2024). Validation of two novel human activity recognition models for typically developing children and children with cerebral palsy. Gait & Posture, 108, 131–137. https://doi.org/10.1016/j.gaitpost.2023.11.013