josefheidler/har_adults_2024-lendt

Lendt et al. 2024 — thigh-worn accelerometry, healthy adults

Thigh-worn accelerometry (SENS motion, 12.5 Hz, ±4 g) with video-labelled activity ground truth in 35 healthy adults (lab) / 32 (free-living), covering both a laboratory protocol and ≈1 h of annotated free-living per subject. Harmonized from the original Zenodo release into per-subject parquet — one row per raw accelerometer sample, with 1 Hz activity labels asof-joined on.

35 participants, 6.9 M samples, 40 % labelled, 61.3 h labeled.

Source

• Paper: Lendt, C., Braun, T., Biallas, B., Froböse, I., & Johansson, P. J. (2024). Thigh-worn accelerometry: A comparative study of two no-code classification methods for identifying physical activity types. International Journal of Behavioral Nutrition and Physical Activity, 21(1), 77. https://doi.org/10.1186/s12966-024-01627-1
• Raw data: Zenodo record 12704412
• License: CC-BY-4.0

Protocol

• Participants: 38 recruited (30.8 ± 9.6 y, 53 % female); 35 included in lab, 32 in free-living — see Harmonization notes.
• Sensor: SENS motion (SENS Innovation, Denmark). Triaxial accelerometer, ±4 g, 12.5 Hz fixed. Adhesive patch on the right lateral thigh, 10 cm above the lateral epicondyle (harmonized data transformed to front-thigh frame — see Axis orientation).
• Lab: ≈46 min per subject across 14 sub-conditions of 3–5 min each — 6 activity types: stand; sit; lie[supine], lie[side], lie[prone]; treadmill walk at 0.5 / 0.8 / 1.2 m·s⁻¹; treadmill run at 1.8 / 2.3 / 2.8 m·s⁻¹; cycle ergometer bicycle at 50 W/40 rpm, 75 W/60 rpm, 100 W/80 rpm. Transitions and waiting periods between sub-conditions are unlabelled.
• Free-living: ≈60 min per subject of unrestricted activity (cycling encouraged), video-annotated throughout.
• Cohort-wide labelled reference: 61.3 h (26.7 h lab + 34.6 h free-living).

Ground-truth provenance

• Lab: timestamps logged manually via a custom R Shiny app at activity onsets.
• Free-living: chest-mounted GoPro Hero 4 Black (30 fps, 1920×1080), frame-by-frame annotation in ELAN v6.4 by a single rater. Inter-rater Kappa 0.95 on 5 double-coded videos.
• Reference resolution: 1 s (start/end rounded).
• Sensor–video sync: 3 heel drops separated by 15 s rest. Lab sync via sensor flip.

Schema

All timestamps are tz-aware Europe/Berlin (recording-site local time).

column	dtype	notes
subject	string	pseudonymized alias, e.g. happy-otter
timestamp	timestamp[ns, Europe/Berlin]	~80 ms cadence (12.5 Hz)
acc_x	float64	hub standard frame (x=up), units = g
acc_y	float64	hub standard frame (y=right), units = g
acc_z	float64	hub standard frame (z=forward), units = g
condition	string (nullable)	laboratory | free-living; null where label is null
label	string (nullable)	activity label; null outside any labeled window
variant	string (nullable)	intensity tier or form variant; null for uncharacterised activities
speed_kph	float64 (nullable)	treadmill speed in km/h (walk/run lab stages only)
power_w	float64 (nullable)	cycling power in W (bicycle lab stages only)

Example rows:

subject      timestamp                   condition    label     variant           speed_kph  power_w
happy-otter  2022-06-01 09:14:32+02:00   laboratory   stand     null              null       null
happy-otter  2022-06-01 09:22:10+02:00   laboratory   walk      slow              1.8        null
happy-otter  2022-06-01 09:35:00+02:00   laboratory   bicycle   fast              null       100.0
happy-otter  2022-06-01 09:55:12+02:00   laboratory   lie       supine            null       null
happy-otter  2022-06-01 10:02:15+02:00   free-living  walk      null              null       null
happy-otter  2022-06-01 10:15:30+02:00   free-living  bicycle   pedalling-seated  null       null
happy-otter  2022-06-01 09:13:01+02:00   null         null      null              null       null

Label vocabulary

label and variant are stored in separate columns. variant is null for activities with no controlled intensity or characterised form.

label values (8): bicycle, lie, run, shuffle, sit, stairs, stand, walk.

Label glossary

Full meaning of every (label, variant) combination in the dataset:

label	variant	condition	meaning	labeled min
bicycle	slow	laboratory	Cycle ergometer 50 W / 40 rpm	104.5
bicycle	moderate	laboratory	Cycle ergometer 75 W / 60 rpm	104.7
bicycle	fast	laboratory	Cycle ergometer 100 W / 80 rpm	104.9
bicycle	pedalling-seated	free-living	Cycling seated on saddle, actively pedalling	237.2
bicycle	coasting	free-living	Cycling without pedalling (downhill / momentum); seated/standing posture not distinguished in free-living annotation	49.4
bicycle	pedalling-standing	free-living	Cycling while standing on pedals	10.3
lie	supine	laboratory	Lying on back, face up	104.9
lie	side	laboratory	Lying on side	104.9
lie	prone	laboratory	Lying face down	104.9
lie	—	free-living	Lying; specific posture not recorded	41.9
run	slow	laboratory	Treadmill 6.48 km/h (1.8 m/s)	104.9
run	moderate	laboratory	Treadmill 8.28 km/h (2.3 m/s)	104.9
run	fast	laboratory	Treadmill 10.08 km/h (2.8 m/s)	101.9
run	—	free-living	Running; speed uncontrolled	134.9
shuffle	—	free-living	Standing with small continuous body movements (dynamic standing)	76.0
sit	—	laboratory	Sedentary; seated at rest	174.8
sit	—	free-living	Sedentary; seated at rest	722.8
stairs	—	free-living	Stair climbing	13.1
stand	—	laboratory	Upright, stationary standing	174.7
stand	—	free-living	Upright, stationary standing	220.1
walk	slow	laboratory	Treadmill 1.8 km/h (0.5 m/s)	104.9
walk	moderate	laboratory	Treadmill 2.88 km/h (0.8 m/s)	104.9
walk	fast	laboratory	Treadmill 4.32 km/h (1.2 m/s)	104.9
walk	—	free-living	Walking; speed uncontrolled	569.3

Total labeled: 3,680 min / 61.3 h across 35 participants (lab) / 32 (free-living).

Coverage

Every raw accelerometer sample is kept, even when no ground-truth label covers it — typically in-lab transitions between sub-conditions or pre/post-protocol margins. Each sample carries the most recent ground-truth label whose 1 Hz timestamp falls at or before the sample's timestamp, within 1 s; samples outside any labelled window have label = null and condition = null.

Cohort-wide: 6.9 M samples, 40 % (2.76 M) labelled. Per-subject labelled fraction ranges from 17 % to 75 %. Filter with df[df['label'].notna()] to keep only labelled rows.

Axis orientation

Accelerometer values are not raw SENS motion output. They are transformed into the hub standard axis convention, shared across all datasets on this hub:

• acc_x runs along the thigh toward the head — reads +1 g when the person stands upright.
• acc_y points to the person's right — positive when tilting right, negative when tilting left.
• acc_z points forward — positive when leaning or stepping forward, negative when leaning backward.
• At rest standing upright: acc_x ≈ +1 g, acc_y ≈ 0, acc_z ≈ 0.

Harmonized accelerometer data is presented in a front-thigh frame — as if the sensor were mounted on the anterior surface of the thigh. The physical sensor (SENS motion) was originally worn on the right lateral thigh (10 cm above the lateral epicondyle); two corrections transform the native lateral-thigh axes to the front-thigh convention:

Native lateral-thigh axes (standing upright): X=down, Y=forward (anterior), Z=left (medial).

1. Negate x — acc_x ← -acc_x. X=down → x=up.
2. Rotate 90° about x — (acc_y, acc_z) ← (-acc_z, acc_y). Y=forward (anterior), Z=left (medial) → y=right, z=forward. This rotation accounts for the ~90° angular difference between lateral and anterior thigh placement.

The result is identical to hub-standard front-thigh data (consistent with the HARTH thigh sensor and all other datasets on this hub).

For raw SENS-native (lateral-thigh) values, ingest the upstream Zenodo record directly (record 12704412).

Harmonization notes

Participant exclusions: 35 of 38 included — 3 dropped for missing upstream annotations; a further 3 dropped from free-living for sensor detachment / poor video (free-living n = 32).

Per-sample exclusions:

• merry-ocelot: run[fast] segments dropped (author-flagged invalid).

clever-lynx's free-living annotations are merged from two upstream folders (both windows from the same continuous recording).

Use

Intended for human activity recognition (HAR) from thigh-worn accelerometry — predicting label (8-class) from the acc_x/y/z signal. The original paper collapses these into 5 classes: sedentary, standing, walking, running, cycling.

The standard evaluation protocol is leave-one-subject-out (LOSO) cross-validation: 35 folds for lab data, 32 for free-living (3 subjects have lab data only). Avoid random row-level splits — they leak temporal and subject-level context across folds.

Loading

Each subject is stored as a separate parquet file under harmonized/. Filenames are pseudonymized aliases (happy-otter.parquet, clever-lynx.parquet, …). Every row carries a subject column with the same alias, so identity is preserved when loading all files together.

Load all subjects into one table:

from datasets import load_dataset

ds = load_dataset("josefheidler/har_adults_2024-lendt")
df = ds["train"].to_pandas()

Load a single subject with pandas:

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/josefheidler/har_adults_2024-lendt/harmonized/happy-otter.parquet"
)

Load all subjects individually (preserving identity):

import pandas as pd
from huggingface_hub import HfFileSystem

fs = HfFileSystem()
files = fs.glob("datasets/josefheidler/har_adults_2024-lendt/harmonized/*.parquet")
dfs = {
    f.split("/")[-1].replace(".parquet", ""): pd.read_parquet(f"hf://{f}")
    for f in files
}

Citation

Lendt, C., Braun, T., Biallas, B., Froböse, I., & Johansson, P. J. (2024). Thigh-worn accelerometry: A comparative study of two no-code classification methods for identifying physical activity types. International Journal of Behavioral Nutrition and Physical Activity, 21(1), 77. https://doi.org/10.1186/s12966-024-01627-1