Best wearable datasets for human activity recognition
Most activity-recognition benchmarks are small lab datasets recorded with phones or body-worn IMUs: UCI-HAR (30 volunteers, waist phone), WISDM (36 users, phone accelerometer), PAMAP2 (9 subjects, 3 IMUs plus heart rate), MHEALTH (10 subjects, 3 sensors plus ECG), USC-HAD (14 subjects, hip IMU), OPPORTUNITY (4 users in a sensor-rich kitchen) and HHAR (9 users, 12 phones and watches). CAPTURE-24 is the main large free-living option, with 151 participants wearing a wrist accelerometer for about 24 hours each. Few of these include PPG, so they mostly serve the motion side of wearable models. Evaluation choices dominate reported scores: on the same four datasets, Yuan et al. (2024) found a self-supervised network pre-trained on UK Biobank wrist data beat a random-forest baseline under subject-wise splits (macro F1 0.726 vs 0.694 on CAPTURE-24, 0.595 vs 0.416 on OPPORTUNITY), while near-perfect scores in other papers often come from splits that share subjects between training and test. The HAR leaderboard lists every verified result by dataset and metric.
Top picks and when to use each
151 participants wore an Axivity AX3 wrist accelerometer for about 24 hours, with wearable-camera and diary labels for more than 2,500 hours; the usual training set for wrist HAR models.
Access: Open (CC BY, Oxford ORA)Original host9 subjects, 3 Colibri IMUs (wrist, chest, ankle) at 100 Hz and a heart-rate monitor at about 9 Hz, 18 activities.
Access: Open (CC BY 4.0, UCI)Original host- #3UCI Human Activity Recognition Using Smartphones (UCI-HAR)Best for: Quick smartphone baseline with an official split
30 volunteers, waist-worn Galaxy S II accelerometer and gyroscope at 50 Hz, 2.56 s windows, 561 features, official 21/9 subject split.
Access: Open (CC BY 4.0, UCI)Original host 4 users × 6 runs with 7 IMUs, 12 accelerometers and object and ambient sensors at about 30 Hz; locomotion, 17 gestures and high-level activity labels.
Access: Open (CC BY 4.0, UCI)Original host15 subjects, 8 activities, acceleration and other phone and watch sensors recorded simultaneously at chest, forearm, head, shin, thigh, upper arm and waist.
Access: Free download (licence not stated)Original host10 volunteers, 12 activities, Shimmer2 sensors on chest, right wrist and left ankle at 50 Hz, with chest ECG.
Access: Open (CC BY 4.0, UCI)Original host9 users with 4 smartwatches and 8 smartphones recording accelerometer and gyroscope at each device's own rate, plus a still-device experiment.
Access: Open (CC BY 4.0, UCI)Original host18 adults aged 70–95 with thigh and lower-back Axivity AX3 accelerometers, video-annotated semi-structured free living.
Access: Open (CC BY 4.0, UCI)Original host14 subjects, 12 activities including walking left/right and elevator rides, MotionNode at 100 Hz, 5 trials each.
Access: Free download (licence not stated)Original host36 users, 6 activities, 20 Hz phone accelerometer, 1.1 million raw samples; parse carefully.
Access: Free download (licence not stated)Original host
Comparison table
20 datasets. Swipe sideways on mobile. Values come from each dataset record; open a dataset for sources. Popularity is cached at build time: OpenAlex citations of the source paper (or the dataset DOI when no paper is linked), results on ChatPPG leaderboards, and GitHub stars or Zenodo / Hugging Face downloads where the host reports them. Hover for the source; a dash means unknown.
| Signals | Sampling | Ground truth | Licence | Access | |||
|---|---|---|---|---|---|---|---|
| capture-24 | 151 | accel | 100Hz | activity, SQI | CC BY 4.0 | Open (CC BY, Oxford ORA) | 60 cited9 results |
| PAMAP2 | 9 | accel, gyro, magnetometer, skin temp, HR | 100 Hz, 9 Hz | HR, activity | CC BY 4.0 | Open (CC BY 4.0, UCI) | 1,444 cited12 results |
| UCI Human Activity Recognition Using Smartphones (UCI-HAR) | 30 | IMU | — | — | Creative Commons Attribution 4.0 International (CC… | Open (CC BY 4.0, UCI) | 77 cited3 results |
| opportunity | 4 | accel | 30Hz | activity | Creative Commons Attribution 4.0 International (CC… | Open (CC BY 4.0, UCI) | 788 cited6 results |
| RealWorld HAR | 15 | accel, gyro, magnetometer, audio, gps | 50 Hz | activity | See host | Free download (licence not stated) | 368 cited4 results |
| mhealth | 10 | accel | 50Hz | activity | Creative Commons Attribution 4.0 International (CC… | Open (CC BY 4.0, UCI) | 580 cited9 results |
| HHAR | 9 | accel | — | activity | Creative Commons Attribution 4.0 International (CC… | Open (CC BY 4.0, UCI) | 712 cited1 result |
| HAR70+ | 18 | accel | — | activity | Creative Commons Attribution 4.0 International (CC… | Open (CC BY 4.0, UCI) | 70 cited |
| USC-HAD | 14 | accel | — | activity | See host | Free download (licence not stated) | 554 cited3 results |
| WISDM AR v1.1 | 36 | accel | 20 Hz | activity | See host | Free download (licence not stated) | 2,902 cited5 results |
| HARTH | 22 | accel | — | activity | See host | See host | — |
| PAAWS R1 | 252 | accel | — | activity | See host | See host | — |
| daily-and-sports-activities | 8 | accel | 25Hz | activity | See host | See host | 13 cited |
| WEAR | 22 | accel | — | activity | See host | See host | — |
| REALDISP | 17 | accel | — | activity | See host | See host | — |
| OpenWatch | 50 | accel, gyro, PPG | — | activity | See host | See host | 0 cited |
| ScientISST MOVE: Annotated Wearable Multimodal Biosignals recorded during Everyday Life Activities in Naturalistic Environments | 17 | PPG, ECG, accel, EDA | 4Hz, 500Hz | — | See host | See host | 3 cited |
| DEO | 90 | accel, gyro, magnetometer | 100 Hz | activity | See host | See host | 1 result |
| μDPad | 65 | PPG, accel, gyro | 112.3 Hz, 112.35 Hz | — | See host | See host | 0 cited148 downloads |
| Skoda mini checkpoint | — | accel | — | activity | See host | See host | — |
Benchmark results
Reported results from papers in the ChatPPG hub. Protocols differ, so only compare rows on the same dataset and metric.
| Dataset | Metric | Method | Result | Protocol |
|---|---|---|---|---|
| capture-24 | Cohen’s κ (e.g. sleep staging) ↑ | HARNet / OxWearables SSL | 0.86 | same 4-class intensity protocol as above; RF+HMM baseline 0.800 |
| capture-24 | Cohen’s κ (e.g. sleep staging) ↑ | SSL ResNet, UK Biobank (Yuan) | 0.737 | Wrist accelerometer resampled to 30 Hz, 10 s windows; subject-wise 5-fold CV (leave-one-subject-out when fewer than 10 s… |
| capture-24 | Cohen’s κ (e.g. sleep staging) ↑ | ResNet from scratch (Yuan) | 0.703 | Wrist accelerometer resampled to 30 Hz, 10 s windows; subject-wise 5-fold CV (leave-one-subject-out when fewer than 10 s… |
| capture-24 | Cohen’s κ (e.g. sleep staging) ↑ | Random forest (Yuan) | 0.683 | Wrist accelerometer resampled to 30 Hz, 10 s windows; subject-wise 5-fold CV (leave-one-subject-out when fewer than 10 s… |
| capture-24 | Macro-F1 on HAR activities ↑ | ActiNet | 0.818 | Sleep/sedentary/light/MVPA (Walmsley), 30-s @30 Hz wrist ACC, 151 pts, nested 5-fold group CV; κ 0.860; RF+HMM 0.764. |
| capture-24 | Macro-F1 on HAR activities ↑ | SSL ResNet, UK Biobank (Yuan) | 0.726 | Wrist accelerometer resampled to 30 Hz, 10 s windows; subject-wise 5-fold CV (leave-one-subject-out when fewer than 10 s… |
| capture-24 | Macro-F1 on HAR activities ↑ | ResNet from scratch (Yuan) | 0.708 | Wrist accelerometer resampled to 30 Hz, 10 s windows; subject-wise 5-fold CV (leave-one-subject-out when fewer than 10 s… |
| capture-24 | Macro-F1 on HAR activities ↑ | Random forest (Yuan) | 0.694 | Wrist accelerometer resampled to 30 Hz, 10 s windows; subject-wise 5-fold CV (leave-one-subject-out when fewer than 10 s… |
| capture-24 | Macro-F1 on HAR activities ↑ | An open-source deep learning model for classifyi | 0.661 | Sedentary/light/MVPA (no sleep), 10-s windows, Capture-24 downsampled 100→12.5 Hz, trained on CP+HP lab + Capture-24 ("A… |
| DEO | Balanced accuracy (HAR / class-imbalance) ↑ | Diffusion-Based Synthetic Data Pretraining for E | 90.6 | CABiGRU pre-trained on diffusion-generated smartwatch IMU windows, fine-tuned on real data |
| HHAR | Macro-F1 on HAR activities ↑ | InrtlWrist | 0.621 probe / 0.603 FT | IMU-Bench task card: 6 classes, 2.56 s, subject-disjoint; LIMU-BERT/SimMTM/FOCAL/RelCon/SSL-Wearables retrained on the s… |
| mhealth | Accuracy ↑ | Random forest (Yang) | 95.09 | Leave-one-subject-out; random-forest-based strategy; the paper reports an average F1 and accuracy in %. |
| mhealth | F1 on HAR activities (averaging not stated) ↑ | MCAE (Azadi) | 99.58 | Held-out test subjects (UCI-HAR uses the official 21/9 subject split); F1 in %, averaging (macro or weighted) not stated… |
| mhealth | F1 on HAR activities (averaging not stated) ↑ | Single-task baseline (Azadi) | 99.23 | Held-out test subjects (UCI-HAR uses the official 21/9 subject split); F1 in %, averaging (macro or weighted) not stated… |
| mhealth | F1 on HAR activities (averaging not stated) ↑ | CAE (Azadi) | 97.33 | Held-out test subjects (UCI-HAR uses the official 21/9 subject split); F1 in %, averaging (macro or weighted) not stated… |
| mhealth | F1 on HAR activities (averaging not stated) ↑ | Random forest (Yang) | 94.49 | Leave-one-subject-out; random-forest-based strategy; the paper reports an average F1 and accuracy in %. |
| mhealth | Macro-F1 on HAR activities ↑ | CHEAR | 88.62 (%) | Cross-subject; prior-method numbers in the paper are copied from ZARA (not re-run) → same-family as the ZARA row (86.1, … |
| mhealth | Macro-F1 on HAR activities ↑ | ZARA | 86.1 (%) | same ZARA protocol, 12 classes. |
| mhealth | Macro-F1 on HAR activities ↑ | OntoAR-HAR | 84.6 ± 2.6 (%) | Mean of ten random samples (Table II); ZARA row copied in-table; 2026-10-14 |
| mhealth | Macro-F1 on HAR activities ↑ | InrtlWrist | 68.6 probe / 72.4 FT (%) | IMU-Bench subject-disjoint 2.56 s, 12 classes, wrist channels only — different protocol from ZARA-family rows; not ranke… |
| opportunity | Macro-F1 on HAR activities ↑ | ZARA | 92.5 (%) | ZARA protocol: knowledge base from seen subjects, class-balanced held-out-subject test, 4 classes (Table 1). |
| opportunity | Macro-F1 on HAR activities ↑ | OntoAR-HAR | 91.5 ± 1.6 (%) | Mean of ten samples; 2026-10-14 |
| opportunity | Macro-F1 on HAR activities ↑ | CHEAR | 85.13 (%) | ZARA 92.5 in the same table; 2026-10-14 |
| opportunity | Macro-F1 on HAR activities ↑ | SSL ResNet, UK Biobank (Yuan) | 59.5 | Wrist accelerometer resampled to 30 Hz, 10 s windows; subject-wise 5-fold CV (leave-one-subject-out when fewer than 10 s… |
| opportunity | Macro-F1 on HAR activities ↑ | Random forest (Yuan) | 41.6 | Wrist accelerometer resampled to 30 Hz, 10 s windows; subject-wise 5-fold CV (leave-one-subject-out when fewer than 10 s… |
| PAMAP2 | Accuracy ↑ | Random forest (Yang) | 92.21 | Leave-one-subject-out; random-forest-based strategy; the paper reports an average F1 and accuracy in %. |
| PAMAP2 | F1 on HAR activities (averaging not stated) ↑ | MCAE (Azadi) | 94.88 | Held-out test subjects (UCI-HAR uses the official 21/9 subject split); F1 in %, averaging (macro or weighted) not stated… |
| PAMAP2 | F1 on HAR activities (averaging not stated) ↑ | Single-task baseline (Azadi) | 92.66 | Held-out test subjects (UCI-HAR uses the official 21/9 subject split); F1 in %, averaging (macro or weighted) not stated… |
| PAMAP2 | F1 on HAR activities (averaging not stated) ↑ | Random forest (Yang) | 91.64 | Leave-one-subject-out; random-forest-based strategy; the paper reports an average F1 and accuracy in %. |
| PAMAP2 | F1 on HAR activities (averaging not stated) ↑ | CAE (Azadi) | 91.0 | Held-out test subjects (UCI-HAR uses the official 21/9 subject split); F1 in %, averaging (macro or weighted) not stated… |
| PAMAP2 | Macro-F1 on HAR activities ↑ | OntoAR-HAR | 87.3 ± 1.1 (%) | Mean of ten samples; 2026-10-14 |
| PAMAP2 | Macro-F1 on HAR activities ↑ | CHEAR | 83.43 (%) | vs ZARA 76.9 in the same table; not ranked against LOSO MorphCL/HARNet; 2026-10-14 |
| PAMAP2 | Macro-F1 on HAR activities ↑ | SSL ResNet, UK Biobank (Yuan) | 78.9 | Wrist accelerometer resampled to 30 Hz, 10 s windows; subject-wise 5-fold CV (leave-one-subject-out when fewer than 10 s… |
| PAMAP2 | Macro-F1 on HAR activities ↑ | ZARA | 76.9 (%) | Held-out-subject class-balanced split, 12 classes — different protocol from the LOSO MorphCL/HARNet row; not ranked agai… |
| PAMAP2 | Macro-F1 on HAR activities ↑ | Random forest (Yuan) | 75.3 | Wrist accelerometer resampled to 30 Hz, 10 s windows; subject-wise 5-fold CV (leave-one-subject-out when fewer than 10 s… |
| RealWorld HAR | Macro-F1 on HAR activities ↑ | SSL ResNet, UK Biobank (Yuan) | 0.792 | Wrist accelerometer resampled to 30 Hz, 10 s windows; subject-wise 5-fold CV (leave-one-subject-out when fewer than 10 s… |
| RealWorld HAR | Macro-F1 on HAR activities ↑ | Random forest (Yuan) | 0.731 | Wrist accelerometer resampled to 30 Hz, 10 s windows; subject-wise 5-fold CV (leave-one-subject-out when fewer than 10 s… |
| RealWorld HAR | Macro-F1 on HAR activities ↑ | ResNet from scratch (Yuan) | 0.705 | Wrist accelerometer resampled to 30 Hz, 10 s windows; subject-wise 5-fold CV (leave-one-subject-out when fewer than 10 s… |
| RealWorld HAR | Macro-F1 on HAR activities ↑ | ActiNet | 0.661 | Deployed CAPTURE-24 model, wrist ACC; RF+HMM 0.531. |
| UCI Human Activity Recognition Using Smartphones (UCI-HAR) | F1 on HAR activities (averaging not stated) ↑ | MCAE (Azadi) | 98.55 | Held-out test subjects (UCI-HAR uses the official 21/9 subject split); F1 in %, averaging (macro or weighted) not stated… |
| UCI Human Activity Recognition Using Smartphones (UCI-HAR) | F1 on HAR activities (averaging not stated) ↑ | Single-task baseline (Azadi) | 97.53 | Held-out test subjects (UCI-HAR uses the official 21/9 subject split); F1 in %, averaging (macro or weighted) not stated… |
| UCI Human Activity Recognition Using Smartphones (UCI-HAR) | F1 on HAR activities (averaging not stated) ↑ | CAE (Azadi) | 94.61 | Held-out test subjects (UCI-HAR uses the official 21/9 subject split); F1 in %, averaging (macro or weighted) not stated… |
| USC-HAD | F1 on HAR activities (averaging not stated) ↑ | MCAE (Azadi) | 83.9 | Held-out test subjects (UCI-HAR uses the official 21/9 subject split); F1 in %, averaging (macro or weighted) not stated… |
| USC-HAD | F1 on HAR activities (averaging not stated) ↑ | Single-task baseline (Azadi) | 76.17 | Held-out test subjects (UCI-HAR uses the official 21/9 subject split); F1 in %, averaging (macro or weighted) not stated… |
| USC-HAD | F1 on HAR activities (averaging not stated) ↑ | CAE (Azadi) | 75.69 | Held-out test subjects (UCI-HAR uses the official 21/9 subject split); F1 in %, averaging (macro or weighted) not stated… |
| WISDM AR v1.1 | Accuracy ↑ | MLP (Kwapisz) | 91.7 | WISDM v1.1 (original release), 29 users, phone in front pocket, 10 s windows, 43 features; 10-fold cross-validation over… |
| WISDM AR v1.1 | Accuracy ↑ | J48 (Kwapisz) | 85.1 | WISDM v1.1 (original release), 29 users, phone in front pocket, 10 s windows, 43 features; 10-fold cross-validation over… |
| WISDM AR v1.1 | Accuracy ↑ | Logistic regression (Kwapisz) | 78.1 | WISDM v1.1 (original release), 29 users, phone in front pocket, 10 s windows, 43 features; 10-fold cross-validation over… |
| WISDM AR v1.1 | Accuracy ↑ | Majority class (Kwapisz) | 37.2 | WISDM v1.1 (original release), 29 users, phone in front pocket, 10 s windows, 43 features; 10-fold cross-validation over… |
| WISDM AR v1.1 | Macro-F1 on HAR activities ↑ | ActiNet | 0.741 | Deployed CAPTURE-24 model, labels mapped to intensity; RF+HMM 0.679. |
Preprocessing pitfalls
- Random window splits leak subjects and overlapping windows into the test set. Use subject-wise cross-validation or leave-one-subject-out, and say which.
- Macro and weighted F1 can differ by many points on imbalanced datasets such as OPPORTUNITY and WISDM; always state which you report.
- Whether the null class is included changes OPPORTUNITY and MHEALTH scores substantially.
- Sensor position matters: a model trained on waist or hip data (UCI-HAR, USC-HAD) does not transfer directly to the wrist.
- Resampling, window length and label mapping differ across papers; Yuan et al. (2024) resampled everything to 30 Hz with 10 s windows, which is not the original UCI or PAMAP2 setup.
- Several popular datasets (WISDM, USC-HAD, RealWorld) do not state a licence on the host page; check before commercial use or redistribution.
FAQ
What is the best dataset for human activity recognition?
For free-living wrist data, CAPTURE-24 (151 participants, about 24 hours each). For controlled multi-sensor benchmarks, PAMAP2 and OPPORTUNITY. For a quick smartphone baseline with an official subject split, UCI-HAR.
Which HAR datasets include heart rate or ECG?
PAMAP2 includes a heart-rate monitor at about 9 Hz and MHEALTH includes 2-lead chest ECG at 50 Hz. Some newer smartwatch datasets in the hub, such as ScientISST MOVE and OpenWatch, add PPG.
What accuracy do models reach on PAMAP2 and CAPTURE-24?
Under subject-wise 5-fold or leave-one-subject-out evaluation at 30 Hz, Yuan et al. (2024) reported macro F1 0.789 on PAMAP2 and 0.726 on CAPTURE-24 for a self-supervised ResNet, against 0.753 and 0.694 for a random forest. Higher figures elsewhere usually come from different splits or label sets.
Is there a HAR leaderboard?
Yes. The ChatPPG HAR leaderboard ranks results only within the same dataset and metric, with the evaluation protocol and a source link on every row.