Public data where walking was measured at the wrist.

Gyroscope and magnetometer data exist only in the separate RawData.zip; the synchronised release (9 columns, wrist/hip/ankle) is accelerometer-only, and both streams run at ~15 Hz. Annotations distinguish true steps from 'shifts' (pivots, shuffles, first/last stride), so step-count evaluations that ignore shifts will mis-score; note cecas.clemson.edu serves over plain HTTP only (broken HTTPS chain), with the Google Sites page carrying the CC BY 4.0 dataset license.

~60,800 individual step events manually annotated frame-by-frame from synchronised video of the feet, labelled left/right and separated from non-step 'shifts'.

One of the very few datasets pairing a wrist accelerometer with hardware-measured (FSR) per-step gait events across treadmill, indoor and outdoor conditions; the wrist sensor is left-wrist only and there is no gyroscope. Download links are emailed after signing a release agreement, no subject appears in every condition (11 indoor vs 9 outdoor), and the right-ankle sensor is mounted in an arbitrary orientation.

Heel-strike and toe-off events from piezo-electric force-sensitive resistors at the ends of the shoe sole, recorded synchronously with the accelerometers.

Among the strongest reference annotations in this index (real-world stride-level ground truth in impaired cohorts), and the wrist unit is a full INDIP magneto-inertial module (accelerometer + gyroscope, 100 Hz) present for 105 of 108 participants — but only in v1.0.2, so earlier records such as zenodo.org/records/13987963 contain no wrist data at all. The record's licence is internally contradictory (description text CC BY-NC-ND 4.0 vs. metadata CC BY 4.0), and all turning parameters were removed from the reference as unvalidated, so there is no turning ground truth.

Marker-based stereophotogrammetric motion capture for the in-lab tests plus the INDIP multi-sensor reference system (two plantar pressure insoles, three IMUs, two distance sensors) for lab and free-living, giving stride-level gait events and spatio-temporal parameters at micro-walking-bout and continuous-walking-period level.

The only public pairing of a consumer smartwatch with force-plate ground truth, so contact events come from measured force rather than from another IMU. Small and laboratory-bound. Primary sources disagree on the licence — the canonical Zenodo deposit says CC0, so confirm before relying on it commercially.

AMTI OR6-7 force plate at 1000 Hz, time-aligned to the watch IMU; vertical ground reaction force is the primary reference signal.

The repeat structure is what makes it valuable: up to 18 four-minute walking trials per person across two days at 200 Hz, which is the only public source for within- and between-day wrist repeatability at this scale. Catches: gait events are algorithmic rather than human-annotated, there is no channel literally named wrist (forearm and hand bracket a watch position), signals carry a Noraxon 5 Hz anti-wobble filter, and the collection is 50 GB. CC0 applies to the DATA; the descriptor article is CC BY 4.0.

Per-sample stance/swing contact signals and per-step spatiotemporal parameters, both computed by Noraxon software from the FOOT IMUs — not an independent reference.

Largest open free-living dataset with video-annotated per-step ground truth at the wrist, and the training set behind the OxWearables 'stepcount' model. Axivity AX3 loggers record acceleration only (no gyroscope), with a paired 100 Hz and 25 Hz device at each of the wrist and hip.

Per-step heel-strike annotations transcribed from synchronised foot-facing action-camera video worn at the beltline.

Rare per-step, per-foot manual labels time-synchronised with a wrist device (Apple Watch Series 10), which makes it directly usable for step-event work. Catch: only 11 of the 16 enrolled participants have signal data (P12 has labels but no data, P13–16 are empty), total recording is ~48 min, and all devices are delivered resampled onto one ~100 Hz timeline with earbud channels forward-filled; wrist side is unstated.

Manual per-step annotations giving start/end sample indices with gait type and foot side (8 classes); the companion paper cites Xsens motion capture as reference but no mocap data is distributed.

Carries wrist gyroscope AND gait-event ground truth, which is the rarest combination in this index — but NC-ND blocks both commercial use and redistribution of derivatives. Listed so the trade-off is visible rather than rediscovered.

INDIP wearable multi-sensor reference (foot-mounted magneto-inertial units plus additional sensors) giving gait events.

Open on PhysioNet with no credentialing, and the upper-limb units give six-axis data at 60 Hz against a full-body reference. Small cohort, and the walker context means the arms are not swinging freely — a real limitation for arm-swing-driven methods.

Xsens MTw Awinda inertial motion capture (MVN BIOMECH) providing full-body segment positions and orientations as the stated kinematic reference, alongside two depth cameras.

Rare pairing of a consumer smartwatch (Galaxy Watch 5, 50 Hz accelerometer and gyroscope on the left or preferred wrist) with a full-body mocap gait-speed reference, plus a real-world COPD patient cohort; the paired smart shoes instrument both feet and the phone has no gyroscope. Catch: the linked Sensors 2024 descriptor documents only version 1 (20 healthy subjects) and the bundled Readme is the sole documentation for the newer subsets, and the free-living COPD recordings carry no per-window speed or activity labels.

Xsens Awinda 17-IMU centre-of-mass horizontal velocity plus walked distances (gait-speed subset), annotated activity start/stop timestamps and known path lengths (activity-recognition and validation subsets), and the average of two clinician-administered 6MWT distances for the COPD cohort.

Rare wrist dataset in which arm-swing amplitude is the manipulated variable, openly downloadable with triaxial accelerometer (m/s^2) and gyroscope (rad/s) at 50 Hz. Catch: the '144 files' in the description are 144 trials stored as cells inside a single 3.9 MB .mat file, and there are no step-event annotations, no participant demographics and no named smartwatch model; the de facto data descriptor is Polvorinos-Fernández et al., BIOSTEC 2025, which the Zenodo record does not link.

Per-trial condition labels only (load level 1–3) plus participant ID; no annotated step events and no external reference system.

Bilateral wrist IMUs with gyroscopes at 128 Hz suit arm-swing modelling, but the reference gait events are IMU-derived rather than from mocap, insoles or manual annotation. No license is declared on the Zenodo record (v1 was deposited as 'Other (Non-Commercial)', later blanked), so the frequently cited CC BY 4.0 covers only the article and reuse terms must be confirmed with the authors.

No independent reference system - gait events and spatiotemporal parameters are computed from the foot IMUs with an error-state Kalman filter plus zero-velocity updates and shipped as 'processed' outputs.

Purpose-built for wrist-to-tibia/shoe signal translation (Kolakowski, Sensors 2024, doi:10.3390/s24010293), so the lower-limb IMU is the target signal and no step or gait-event labels ship with the data. Small and explicitly preliminary: 10 MB, 5 walks, 12 Python-only .pbz2 (pickle) files, wrist in all walks but all three sites simultaneously only in test1 and test5, and no participant count is stated anywhere; the widely quoted "85 min" is the authors' derived training-set size, not the recording duration.

None — the deposit contains no annotation or event files; the concurrently recorded tibia/shoe IMU signals serve as the translation target for the wrist signal, with no optical, pressure, video or manual reference.

Largest clinical gait dataset carrying bilateral wrist IMUs (Xsens MTw Awinda, accel + gyro + magnetometer) alongside instrumented-walkway reference, so wrist signals can be regressed directly against stride parameters. Files live on Synapse and need free registration plus acceptance of the terms; the data descriptor (2026 publication year, data collected earlier) is open at PMC13009270.

ProtoKinetics Zeno pressure-sensing walkway spatiotemporal gait parameters synchronised to all sensors, plus frame-by-frame dual-camera video annotation and clinical scores (MDS-UPDRS, Hoehn & Yahr, medication/DBS status).

Wrist gyroscope exists only for the REAL-PD arm (Moto 360 Sport on the most-affected wrist); CIS-PD's Apple Watch Series 2 is accelerometer-only, and no cohort carries gait or step annotation, so this suits free-living/self-supervised work rather than step validation. Data arrive as 20-minute segments with ordering destroyed so continuous recordings cannot be reconstructed, and the commonly quoted counts (51 enrolled/39 completed CIS-PD, 25 PD + 25 controls REAL-PD) are parent-study enrollment - no controls are in the release.

Self-reported diary labels at 30-minute intervals, released as medication on/off (0-4 in CIS-PD, binary in REAL-PD), dyskinesia and tremor severity; no gait, step-event, walking-bout or video annotation.

Bilateral 100 Hz GENEActiv wrist accelerometry over ~394 participant-days, downloadable from OSF anonymously with no registration, but labels are self-reported at 1-hour granularity, so it supports motor-state and freezing-of-gait work and not stride- or step-event validation. 46 of the 66 participants have deep brain stimulation implanted, which materially affects gait and movement characteristics.

Hourly patient-completed symptom diaries (kinetic ON/OFF 7-point, tremor, binary freezing of gait, falls, medication intake times) plus UPDRS-III in medication-ON and -OFF states; no gait, step or walking-bout annotation.

Wrist IMU with gyroscope at 60 Hz alongside FoG labels, but the reference is clinician-rated episodes and activity classes rather than per-step events, and the 30 fps annotation video is not distributed (only sensor and clinical CSVs are). Volume is small — mean 4.2 min per subject, ~1.5 h total — all recorded in the practically-defined OFF state after levodopa washout, and only one wrist is instrumented with the side varying by participant.

Video-derived annotation by two neurologists of 101 freezing-of-gait episode onsets/offsets with ordinal manifestation labels (shuffling, trembling, akinesia) and seven activity classes, plus subject-level clinical scores (Hoehn & Yahr, MDS-UPDRS III, FOG-Q, MoCA, FES-I, PDQ-8).

The closest external cohort to this project's own input: both wrists at 100 Hz with gyroscope, and frailty labelled in three stages rather than a binary — which is the target where our own binary comparison cannot separate anything and the three-stage one can. 668 participants against our 213. The limit is the task: TUG is supervised and structured, so it can test whether three-stage frailty is readable from wrist six-axis data, not free-living performance. Five sensors (both wrists, both ankles, sacrum); 'lower back' here is the sacrum. Hosted on figshare (article 29315771 v3), not Zenodo as some secondary sources state; the licence was read from the figshare API record rather than the paper. 325 MB across seven files, of which the 292 MB IMU_signals.zip is the raw signal. Verified after download: IMU_signals.zip holds 3,341 CSVs, one per participant per site (left/right wrist, left/right ankle, sacrum), each 100 Hz with Acc XYZ in m/s2, Gyr XYZ in deg/s and roll/pitch/yaw. The 30-column parameter table is entirely left/right FOOT measures, so anything wrist-based has to be built from the raw wrist CSVs. Which of 0/1/2 is robust is not stated in the shipped description.

Three frailty classes assigned by established clinical criteria, with demographic, health and cognitive data attached; the shipped label column splits 342 / 279 / 47, so the frail group is 47 of 668 (7%). Spatiotemporal parameters ship alongside the raw signals but are derived from the ankle sensors by the same pipeline, so they are not an independent reference — and they are foot parameters, not wrist ones.

Qualifies on forearm rather than wrist: the five Shimmer3 units sit on the forearms, shanks and L5, so signals differ from smartwatch data, and the wrist-worn GENEActiv/Pebble data of the same study is the separate 'minimum sensor set' record. There is no backwards-walking task (the dual-task is forward walking while counting backwards aloud), and the CC BY 4.0 belongs to the article, not the dataset, which is governed by Synapse Conditions of Use plus Certified User status.

Limb-specific scores from an MDS-UPDRS-certified clinician: tremor 0-4, dyskinesia 0-4, bradykinesia yes/no (0-4 only for rapid alternating hand movements), plus MDS-UPDRS Part III; no step events, video, motion capture or speed.

Bilateral wrist accelerometry with clinician symptom ratings time-locked to several distinct labelled walking tasks, but accelerometer only, so no wrist angular velocity. The CC BY 4.0 usually quoted covers the article, not the data (Synapse syn20681023 needs a validated, certified account and an Intended Data Use statement); the Spaulding subgroup's five Shimmer3 units are released as the separate 'limb and trunk' record.

Clinician MDS-UPDRS Part III plus per-task tremor severity 0-4, dyskinesia and bradykinesia presence, with task start/end timestamps; no step events, motion capture or measured walking speed.

Rare prodromal-PD cohort with continuous wrist accelerometer + gyroscope wear, but the 100 Hz rate comes only from a secondary analysis paper (Sensors 2022) — neither the PPMI access page nor the two wearable papers publishes a rate, and the earlier "30 Hz for non-PD cohorts" claim is unsupported. Access needs a signed PPMI Data Use Agreement via LONI IDA, Verily's derived-measure algorithms are proprietary, and it is undocumented whether raw gyroscope streams (as opposed to accelerometer) are actually distributed to applicants.

In-person clinical assessment — MDS-UPDRS confirmed for this sub-study, other PPMI scales unverified — plus Verily-derived sleep, vital-sign and physical-activity summaries; no gait-event or stride-level labels.

Xsens wrist IMUs give acceleration, angular velocity, magnetometer and orientation quaternions at 120 Hz in a stroke cohort, but there is no optical mocap, video or walkway reference and no step/gait-event annotation. Wrist data is missing entirely from one of the 20 sessions (imu0011/visit1), and the short name "StrokeGait" is editorial — the record carries no short name and never states the sensor sites or rate (both read off the archive).

Clinical labels only — FAC scores at both visits, two clinician improvement ratings and demographics; the supplied stride parameters are the authors' own foot-IMU-derived outputs, not an independent reference.

Rare six-axis wrist data from an older clinical population, which is the cohort most gait work actually targets. Only 8 of the 27 enrolled participants wore the wrist sensor, and there is no gait reference of any kind.

None for gait — the record positions the data for semi-supervised learning rather than for validation.

Both wrists are instrumented with six-axis data at 100 Hz and the reference is optical motion capture rather than another IMU — an unusually strong combination. Foot markers make it possible to derive contact events even though none are shipped.

Vicon optical motion capture at ~100 Hz, released as 3D marker positions for both hands and both feet.

Six-axis wrist data with participant demographics attached, which is rarer than it should be. At 20 Hz the Nyquist limit is 10 Hz, too low for heel-strike timing; demographics cover only part of the 33 participants.

Per-sample activity class labels only; no event, motion-capture or clinical reference.

Unusually useful for gait because three walking conditions and one running condition carry known treadmill speeds while both arm units sit at the wrists (placement is documented in Altun/Barshan, Pattern Recognition 2010 — the UCI page says only "arms"). Catches: data ships only as pre-cut 5 s segments with no timestamps or step events, 25 Hz limits event timing to 40 ms, and the wrist units are the plus/minus 5 g pair so they can clip during running, jumping and basketball; donated to UCI in 2013.

Activity labels encoded by folder structure over pre-cut 5 s segments, with known treadmill reference speeds for the walking and running trials; no video or optical reference.

Raw watch accelerometer ships as its own archive, so wrist-only use does not require the multi-gigabyte phone streams — but the watch has no gyroscope, labels are self-report, and the cohort is university-age. HTTP-only host: an index or fetcher that upgrades to HTTPS records it as dead. No licence instrument is stated, only a citation condition.

Self-reported labels entered in the study app; no camera, diary or observer reference.

Six-axis wrist data at ~100 Hz with an age range of 23-66, wider than the usual student cohort. The walking is a short Timed-Up-and-Go leg, on the order of twenty minutes of wrist walking across the whole cohort, so it suits transition detection rather than sustained gait. The GitHub licence badge says Apache-2.0, which governs the code, not the data.

Per-sample activity labels produced by frame-level manual inspection of 60 fps video to locate transitions; no step-level annotation.

Sampling-rate and device heterogeneity is the whole point — actual per-device rates deviate from nominal and drift, so it is a rate-robustness stress test rather than a clean gait source and you must not assume a uniform rate for cadence. Gyroscope exists only in the activity-recognition subset (the still-experiment files are accelerometer-only), and the four smartwatches are worn two per arm, so placement is arm/forearm with exact wrist position undocumented.

Per-sample scripted activity labels with a seventh "null" class for unlabelled and transition periods; no step events, treadmill speed, video or motion capture (an Xsens reference was used only in a static zero-g offset test).

Nine-axis at the wrist including a magnetometer, but weak for gait on every other axis: 25 Hz, signals already low-pass filtered and smoothed with no raw release, transitions deliberately deleted, and roughly 6.6 hours total across 16 subjects and 12 activities. Requires a paid IEEE DataPort subscription despite the CC BY licence.

Activity class labels only; no per-step or gait-event annotation.

Left-wrist accelerometer recorded simultaneously with left hip and both ankles under a controlled outdoor walking protocol on four ActiGraph GT3X+ units (accelerometer only — not the gyroscope-carrying GT9X Link). Only activity-type labels ship with the data: any ankle-derived step reference must be computed by the user and is not part of the record, and the ODbL share-alike terms differ from CC BY.

Activity-type label at every time point of data collection (1 = walking, 2 = descending stairs, 3 = ascending stairs, 4 = driving, 77 = clapping, 99 = non-study activity), with no per-step or per-stride event annotation distributed.

Right-wrist Shimmer2 unit carrying accelerometer, gyroscope and magnetometer at 50 Hz, with the side fixed for all subjects (unlike PAMAP2's dominant-arm rule), collected out-of-lab with no constraints on execution style. The wrist unit is documented as "right-lower-arm" in the column list, the chest unit contributes acceleration and 2-lead ECG only (no gyroscope or magnetometer columns), and roughly 1 min per activity across 10 subjects means very little walking data.

Per-sample activity label (column 24; labels 1-12 with 0 as the null class); the UCI page states each session was recorded with a video camera, but whether that video is distributed is unknown.

The gyroscope lives on the lower-arm IMUs (RLA/LLA), not on the wrist units — RWR/LWR are 3-axis accelerometer only, so a wrist-only consumer gets no rate-of-turn. Walking is an explicit label but a minority of each recording (~13% of S1-ADL1) with only 4 subjects, and the docs warn of ~100 ms inter-system and label jitter that matters for step-event timing.

Multi-track manual annotation (4 locomotion modes incl. Walk, 5 high-level activities, 17 mid-level gestures, per-arm action/object primitives), labelled during recording and reviewed by two annotators against video that is not distributed.

Unusually large for a labelled wearable study and the annotation is second-by-second rather than bout level, but the gyroscope exists only in the lab protocol; the free-living arm is accelerometer-only. No licence is stated anywhere on the study site, so reuse terms are unconfirmed.

Second-by-second expert manual annotation from video — a third-person camera in the lab and a chest-mounted camera during waking hours.

The most-used multi-position HAR benchmark that genuinely includes a wrist IMU (3D accelerometer, gyroscope and magnetometer at 100 Hz) with walking, Nordic walking, running and stairs in the activity set. Wrist is on the dominant arm so side varies across subjects, the wrist block is named "IMU hand" (columns 4-20) in the file schema, the per-IMU orientation columns are flagged invalid by the authors, and missing values are NaN.

Per-sample activity label (column 2 of each .dat file), described by the repository as the ground truth and following the scripted collection protocol; no video or optical reference appears in the release.

Best public set for testing how robust a wrist/forearm gait algorithm is to strap rotation and translation, thanks to its three placement scenarios (ideal, self-placed, deliberately displaced; the displaced scenario covers only subjects 2, 5, 15). Catches: the arm units are mid-forearm, not wrist, labels are activity-level only, and recordings are missing for subjects 6 and 13 (self-placement) and nearly all of subject 7 (ideal); rate and placement come from Banos et al., Sensors 2014, not the UCI page.

Activity-level labels annotated online by the experimenter with the CRN Toolbox and later reviewed against video recordings that are not released; no motion capture, treadmill speed or step-event reference.

The same 8 activities are recorded simultaneously at 7 body positions on the same 15 subjects, so forearm signals can be regressed directly against thigh/shin/waist. The landing page states no rate or licence: 50 Hz comes from the readMe bundled in each archive, and the demonstration videos live only on the per-subject pages, not in the ~3.5 GB bulk download.

Activity labels from the scripted recording protocol, plus one demonstration video per activity per subject; no step-event, motion-capture or speed reference.

Wrist smartwatch accelerometer + gyroscope at 50 Hz (m/s^2 and rad/s), but the only labels are gait condition and participant ID, so it suits condition classification rather than step detection. The 72 trials ship as one .mat file (not 72 files, as often reported), and although Zenodo tags the record "Parkinson's Disease" the cohort is entirely healthy.

Per-sample gait-condition label (1=normal, 2=upstairs, 3=downstairs) and participant ID; no step events, mocap, video or clinical scores.

Catalogued as smart-home research, which is why gait searches miss it. The 20 Hz wrist stream suits ambulation classification rather than step timing, and it arrives over connectionless BLE with no delivery guarantee, so the grid is gappy. The 872 test folders are shuffled slices, not participants.

Trained annotators labelled posture, ambulation, activity and location in ELAN against head-mounted 4K reference video; every sequence was annotated by two or three people and interrater kappa is published.

Largest subject count of the classic UCI wrist HAR sets (51) and its deliberate contrast between ambulation and hand-dominated activities is exactly the confounder set a wrist step counter must reject. Two traps: the 20 Hz rate is a nominal OS polling request, not guaranteed, and the raw timestamps are Android nanosecond uptime counters despite the README calling them Unix time — do not treat them as epoch when deriving cadence.

Per-sample activity codes from a scripted 3-minute-per-activity protocol, plus pre-computed 10 s window features; no video or optical reference, but subject IDs are retained for biometrics.

The reference labelled wrist dataset and the training set behind most UK Biobank activity models, with walking as an explicit label — but accelerometer-only, and labels are activity-bout level, so there are no step events, heel strikes or walking speed. The data is CC BY 4.0 while the OxWearables benchmark CODE carries a non-commercial Oxford licence; the two are routinely conflated.

Wearable-camera annotation: a neck-worn Autographer photographed every 20-40 s and trained annotators coded the images to activity codes; sleep came from a two-question diary.

One of very few labelled wrist datasets in adults over 60, with a simultaneous energy-expenditure criterion — but NC-SA blocks commercial use, there is no gyroscope, and the 300 s walking bouts exist for only 25 of the 35 participants. The published collection dates contradict the shipped per-session timestamps, so the year is left unconfirmed.

Researcher-observed activity start and end times imputed onto samples; energy expenditure from simultaneous indirect calorimetry (COSMED).

The largest free-living wrist dataset with a GYROSCOPE at a usable rate, which almost nothing else offers — research-grade wrist loggers are overwhelmingly accelerometer-only. Nothing marks where walking occurs, so it suits pre-training rather than validation. Single 46.7 GB download.

None — the record states no activity labels were collected, the resource being intended for unsupervised and self-supervised learning.

Largest fully open raw wrist-accelerometry population dataset — per-participant tar.bz2 archives on the CDC FTP with no registration, ~1.04 TB compressed for the 2011–2012 half alone; the 2013–2014 half is the separate PAX80_H record, and 247 children aged 3–5 are withheld to the NCHS Research Data Center. Walking is present but nothing marks where it occurs, so it is unusable as a step-detection or gait-parameter validation set; the ActiGraph GT3X+ records triaxial acceleration plus 1 Hz ambient light and no gyroscope, and NCHS use restrictions (statistical use, no re-identification or linkage) still apply.

None — documentation states no activity logs, diaries or records were kept; only per-participant QC flag files marking implausible signal segments (~0.26%) accompany the data.

The reference population-scale wrist dataset and raw data really is released (Axivity AX3 .cwa, ~250 MB per 7-day wear, nominal 100 Hz but true rate fluctuates 94–104 Hz at ±8 g), yet it is accelerometer-only with zero step or gait-event labels — the widely used UKB step models are trained on CAPTURE-24/OxWalk and merely deployed here. Access is by application under the Material Transfer Agreement with bulk-retrieval charges, and was reported suspended as of July 2026 after the Research Analysis Platform closure — verify current status before planning work.

None — no diary, camera or observer annotation accompanies the recordings; the Showcase activity-class fields are algorithm outputs rather than labels.