A Multi-Modal Dataset for Ground Reaction Force Estimation Using Consumer Wearable Sensors

Parvin Ghaffarzadeh; Debarati Chakraborty; Koorosh Aslansefat; Ali Dostan; Yiannis Papadopoulos

eess.SPMar 19, 2026

A Multi-Modal Dataset for Ground Reaction Force Estimation Using Consumer Wearable Sensors

Authors: Parvin Ghaffarzadeh, Debarati Chakraborty, Koorosh Aslansefat, Ali Dostan, Yiannis Papadopoulos

Organizations: Department of Data Science & Artificial Intelligence and Modelling (DAIM), University of Hull, UK · School of Computer Science, University of Hull, UK · School of Science & Technology, Nottingham Trent University, UK

Abstract

This Data Descriptor presents a fully open, multi-modal dataset for estimating vertical ground reaction force (vGRF) from consumer-grade Apple Watch sensors with laboratory force plate ground truth. Ten healthy adults aged 26--41 years performed five activities: walking, jogging, running, heel drops, and step drops, while wearing two Apple Watches positioned at the left wrist and waist. The dataset contains 492 validated trials with time-aligned inertial measurement unit (IMU) recordings (approximately 100 Hz) and force plate vGRF (Force_Z, 1000 Hz). The release includes raw and processed time series, trial-level metadata, quality-control flags, and machine-readable data dictionaries. Trial-level matching manifests link recordings across modalities using stable identifiers. Of the 492 validated trials, 395 are triad-complete, containing wrist, waist, and force plate data, enabling cross-sensor analyses and reproducible model evaluation. Dataset quality is characterised through a three-phase cross-sensor plausibility and consistency framework, repeatability analysis of peak vGRF (intraclass correlation coefficient 0.871--0.990), and systematic checks of force ranges and trial completeness. Monte Carlo sensitivity analysis showed that correlation-based validation metrics were robust to single-sample timing perturbations at the IMU sampling resolution. All data are released under CC BY 4.0, with analysis scripts archived alongside the dataset and mirrored on GitHub. This resource supports reproducible research in wearable biomechanics, benchmarking of machine learning models for vGRF estimation, and investigation of sensor placement effects using widely available consumer wearables.

Figures & tables

Dataset	Year	Sensor Placement	Force Plate	Primary Activities	N	Public
Winter [ 2 ]	1983	Motion capture	Yes	Walking	1	Yes
Moore et al. [ 8 ]	2015	Motion capture	Yes	Walking, perturbations	15	Yes
Fukuchi et al. [ 9 ]	2017	Lower-limb markers	Yes	Running (multiple speeds)	28	Yes
Fukuchi et al. [ 7 ]	2018	Lower-limb markers	Yes	Walking (multiple speeds)	42	Yes
Hu et al. [ 10 ]	2018	Shank, thigh	Yes	Stairs, ramps, level ground	10	Yes
Schreiber & Moissenet [ 12 ]	2019	Lower-limb markers, EMG	Yes	Walking (multiple speeds)	50	Yes

Table 1: Comparison of the present dataset with existing open-source biomechanical datasets combining wearable sensors and ground truth measurements.

Participant	Sex	Age (yr)	Height (cm)	Weight (kg)	BMI (kg/m 2 )	Shoe size	Shoe type
P1	M	35	186.5	95.8	27.5	44	Trainer
P2	F	26	165.5	74.5	27.2	40	Trainer
P3	M	36	172.0	67.8	22.9	41	Trainer
P4	F	32	176.0	68.6	22.1	39	Trainer
P5	M	39	174.0	74.0	24.4	43	Trainer
P6	F	34	162.0	68.2	26.0	37	Trainer

Table 2: Participant demographics (

n=10

Figure 1: Dataset folder structure and organisation. Hierarchical organisation showing 10 participant subfolders (P1–P10), each containing time-aligned IMU sensor files for wrist and waist positions across five activities. Each trial is identified by a unique UUID enabling precise cross-sensor matching. The alignment pipeline produced 598 aligned outputs from 678 matched triplets (pre-curation). The released dataset comprises 492 validated trials documented in trial_manifest.csv (395 triad-complete).

Participant	Walk	Jog	Run	Heel Drop	Step Drop	Total
P1	10	10	7	11	10	48
P2	11	10	10	10	10	51
P3	9	10	10	10	10	49
P4	11	10	10	11	10	52
P5	11	10	11	11	11	54
P6	11	10	10	10	10	51

Table 3: Trial counts per participant and activity (all 492 validated trials).

Participant	Walk	Jog	Run	Heel Drop	Step Drop	Total
P1	7	0	6	11	10	34
P2	9	10	8	7	9	43
P3	9	10	10	10	10	49
P4	11	10	10	11	10	52
P5	11	10	11	0	0	32
P6	9	7	9	8	9	42

Table 4: Trial counts per participant and activity (395 triad-complete trials only).

Figure 2: Trial distribution overview. (A) Per-participant trial counts for all 492 validated trials. (B) Activity-specific trial counts showing triad-complete (all sensors) versus total validated trials. The dataset contains 395 triad-complete trials (80.3%) suitable for comprehensive cross-sensor analyses.

Figure 3: Figure 3. Data collection setup and representative time-aligned signals. Left: (A) Experimental setup showing the waist-worn Apple Watch secured with an elastic belt at the anterior waist (below the navel) and the wrist-worn device on the left wrist; (B) participant standing on the 20 cm step platform used for step-drop trials; (C) MyCoreMotion data collection interface; (D) processing pipeline overview. Right: Representative time-aligned signals showing vertical ground reaction force from the force plate (Force_Z, black), waist acceleration (blue), and wrist acceleration (red). Vertical lines indicate initial contact ( $>$ 50 N). Walking shows a double-peak pattern, running a single higher-amplitude peak, and drop tasks a rapid high-force onset ( $\sim$ 2.5–4.0 $\times$ body weight).

Activity	ICC (95% CI)	CV (%)	N
Walking	0.962 (0.941–0.976)	4.2	96
Jogging	0.981 (0.972–0.987)	3.1	100
Running	0.990 (0.985–0.993)	2.3	97
Heel Drop	0.871 (0.823–0.907)	7.8	100
Step Drop	0.889 (0.847–0.920)	6.9	99

Table 5: Repeatability metrics for peak vGRF.

Characteristic	Description
Subject area	Data Science, Biomedical Engineering, Sports Science
More specific subject area	Ground reaction force estimation using consumer wearable sensors
Type of data	Time series (force plate, IMU accelerometry, gyroscope), metadata (JSON), trial identifiers (UUID)
How data were acquired	AMTI OR6-7 force plate (1000 Hz); Apple Watch Series 5+ accelerometers/gyroscopes ( $\approx$ 100 Hz)
Data format	Raw and processed (CSV files); Python 3.9+ analysis scripts
Parameters for data collection	10 healthy adults (26–41 years, 6 male/4 female), 5 activities, 492 validated trials

Table 6: Dataset overview and key characteristics.

Activity	N	Phase 1 Mean $r$	Phase 3 Mean $r$
Walking	85	0.543	0.430
Jogging	71	0.608	0.582
Running	79	0.624	0.558
Heel Drop	79	0.470	0.294
Step Drop	81	0.442	0.548
Overall	395	0.550	0.486

Table 7: Activity-specific cross-sensor validation performance for Phase 1 and Phase 3 (395 triad-complete trials from primary manifest).

By phase (all 598 trials; mean over comparisons within phase):
Subset	N	Mean $\|\Delta r\|$	P95 $\|\Delta r\|$	Max $\|\Delta r\|$
All comparisons	1,794	0.0111	0.0333	0.1624
$\|r\|\geq 0.2$ comparisons	957	0.0103	0.0306	0.1491
Phase 1 (Waist $\to$ FP)	598	0.0079	—	0.1212
Phase 2 (Wrist $\to$ Waist)	598	0.0144	—	0.1624
Phase 3 (Wrist $\to$ FP)	598	0.0111	—	0.1491

Table 8: Monte Carlo alignment sensitivity summary (598 trials; 1,794 comparisons;

\pm

10ms perturbations). P95 denotes the 95th percentile across comparisons of per-comparison mean

|\Delta r|

. Phase rows summarise mean

|\Delta r|

over comparisons within each phase (598 comparisons per phase).

Figure 4: Time-alignment sensitivity overview ( $|\text{baseline }r|\geq 0.2$ ). (A) Distribution of per-comparison mean $|\Delta r|$ by validation phase. Violin plots show kernel density estimates; box plots show median, interquartile range, and outliers. (B) Relationship between baseline correlation and sensitivity. Vertical dashed lines mark the $|r|=0.2$ threshold; the horizontal dotted line indicates a sensitivity threshold of 0.05.

Figure 5: Empirical cumulative distribution of sensitivity by phase ( $|\text{baseline }r|\geq 0.2$ ). Vertical dashed lines mark thresholds of 0.03 and 0.05.

Figure 6: Threshold exceedance rates by phase ( $|\text{baseline }r|\geq 0.2$ ). Bars show the fraction of comparisons exceeding sensitivity thresholds of 0.03, 0.05, and 0.10.

Figure 7: Mean sensitivity by activity and phase ( $|\text{baseline }r|\geq 0.2$ ). Heatmap shows mean $|\Delta r|$ for each activity-phase combination.

Wearable Sensing

Wearable sensing uses body-worn devices to measure motion, physiology, and interaction for health, activity, and assistive applications.

See all 130 papers

HARMES: A Multi-Modal Dataset for Wearable Human Activity Recognition with Motion, Environmental Sensing and Sound
May 4, 2026Robin Burchard, Pascal-André Brückner, Marius Bock +2Human Activity Recognition Wearable Sensing
University of Siegen, Siegen, Germany · University of Bonn and & Lamarr Institute for Machine Learning and Artificial Intelligence, Bonn, Germany
Selective Correlation Based Knowledge Distillation for Ground Reaction Force Estimation
Apr 27, 2026Eun Som Jeon, Jisoo Lee, Huisu Lim +3Gait Analysis Wearable Sensing
Department of Computer Science and Engineering, Seoul National University of Science and Technology, Seoul, 01811, Republic of Korea · Geometric Media Lab, School of Computing and Augmented Intelligence, Arizona State University, Tempe, 85281, AZ, USA · School for Engineering of Matter, Transport and Energy, Arizona State University, Tempe, 82587, AZ, USA +1
Gravity-guided Contact Dynamics Estimation from 3D Human Motions
Jun 6, 2026Cuong Le, Urs Waldmann, Bastian Wandt +1Contact Force Estimation
Linköping University

May 4, 2026J/K move · Enter open · S save

HARMES: A Multi-Modal Dataset for Wearable Human Activity Recognition with Motion, Environmental Sensing and Sound

Robin Burchard, Pascal-André Brückner, Marius Bock, Juergen Gall+1

University of Siegen, Siegen, Germany · University of Bonn and & Lamarr Institute for Machine Learning and Artificial Intelligence, Bonn, Germany

Human Activity Recognition Wearable Sensing Multimodal Sensor Fusion

Open paper AskPDF arXiv

With each sensing modality exhibiting inherent strengths and limitations, multi-modal approaches for wearable Human Activity Recognition (HAR) are becoming increasingly relevant -- particularly for recognizing Activities of Daily Living (ADLs), where individual modalities often produce ambiguous signals for similar or complex activities. This work introduces HARMES, a multi-modal wearable dataset combining three wrist-recorded modalities: motion sensing via an Inertial Measurement Unit (IMU), atmospheric environmental sensors (humidity, temperature, and pressure), and audio. Collected from 20 participants performing household activities in their own homes, HARMES totals over 80 hours of recorded data, with approximately three hours of labeled activity data per participant across 15 ADL classes. To the best of our knowledge, HARMES is the first dataset to combine this particular sensor trio, and it is nearly six times larger than the previously largest wrist-inertial-acoustic HAR dataset. In an extensive benchmark, we evaluate cross-subject generalization and conduct an ablation study revealing that modality contributions are activity-dependent and can provide complementary value, particularly for activities that are ambiguous from motion data alone. HARMES is freely available at Zenodo, alongside example code for loading the dataset and training models on GitHub.

Participant	Walk	Jog	Run	Heel Drop	Step Drop	Total
P1	10	10	7	11	10	48
P2	11	10	10	10	10	51
P3	9	10	10	10	10	49
P4	11	10	10	11	10	52
P5	11	10	11	11	11	54
P6	11	10	10	10	10	51

Participant	Walk	Jog	Run	Heel Drop	Step Drop	Total
P1	7	0	6	11	10	34
P2	9	10	8	7	9	43
P3	9	10	10	10	10	49
P4	11	10	10	11	10	52
P5	11	10	11	0	0	32
P6	9	7	9	8	9	42

Participant	Walk	Jog	Run	Heel Drop	Step Drop	Total
P1	10	10	7	11	10	48
P2	11	10	10	10	10	51
P3	9	10	10	10	10	49
P4	11	10	10	11	10	52
P5	11	10	11	11	11	54
P6	11	10	10	10	10	51

Participant	Walk	Jog	Run	Heel Drop	Step Drop	Total
P1	7	0	6	11	10	34
P2	9	10	8	7	9	43
P3	9	10	10	10	10	49
P4	11	10	10	11	10	52
P5	11	10	11	0	0	32
P6	9	7	9	8	9	42

Abstract

Wearable Sensing

Explore similar work

HARMES: A Multi-Modal Dataset for Wearable Human Activity Recognition with Motion, Environmental Sensing and Sound

Selective Correlation Based Knowledge Distillation for Ground Reaction Force Estimation

Gravity-guided Contact Dynamics Estimation from 3D Human Motions

Participant	Walk	Jog	Run	Heel Drop	Step Drop	Total
P1	10	10	7	11	10	48
P2	11	10	10	10	10	51
P3	9	10	10	10	10	49
P4	11	10	10	11	10	52
P5	11	10	11	11	11	54
P6	11	10	10	10	10	51

Participant	Walk	Jog	Run	Heel Drop	Step Drop	Total
P1	7	0	6	11	10	34
P2	9	10	8	7	9	43
P3	9	10	10	10	10	49
P4	11	10	10	11	10	52
P5	11	10	11	0	0	32
P6	9	7	9	8	9	42