{"success":true,"data":{"author":"Norwegian University of Science and Technology (Logacjov, Bach, Kongsvold, Bårdstu, Mork); harmonized by Josef Heidler","bucketAccessType":"public","bucketName":"josefheidler/har_adults_2021-harth","bucketPrefix":"","category":"Physiology","cloudProvider":"huggingface","createdAt":"2026-11-22T01:09:18.252Z","description":"HARTH, the Human Activity Recognition Trondheim dataset from the Norwegian University of Science and Technology (NTNU), in its 31-participant v2.0 version: raw triaxial accelerometry from two Axivity AX3 sensors worn on the right thigh and the lower back (about L3) by healthy adults during 1 to 2.5 hours of semi-free-living activity at work or home, plus walking, running and cycling sessions outdoors. Ground truth comes from frame-by-frame annotation of a chest-mounted GoPro video by two experts (Fleiss kappa 0.96) for 12 activity types including walking, running, stair climbing, cycling and postures. The sensors sample at 50 Hz (downsampled from 100 Hz), giving about 8.3 million samples and 46 labelled hours. The setup mirrors the dual-sensor accelerometry of the HUNT4 population study (Trondheim Health Study, about 35,000 people), so the data train physical activity and posture classifiers for wearable epidemiology. This Hugging Face copy harmonizes the NTNU GitHub release into one Parquet file per subject under harmonized/, one row per sample with x, y, z acceleration in g for both sensors and the label; load with Hugging Face datasets or pandas. NTNU publishes the data folder in its harth-ml-experiments repository under the MIT License (v1.2 with 22 subjects is on UCI under CC BY 4.0). Cite Logacjov et al., Sensors 21(23):7853, 2021.","displayTitle":"HARTH (Human Activity Recognition Trondheim) v2.0: Thigh and Lower-Back Accelerometry of 31 Adults with Video-Annotated Activities (NTNU)","huggingFaceUrl":"https://huggingface.co/datasets/josefheidler/har_adults_2021-harth","id":"506a5666-4162-5ed7-964b-f92d0b8f581a","license":"MIT","licenseEvidenceUrl":"https://raw.githubusercontent.com/ntnu-ai-lab/harth-ml-experiments/main/LICENSE","objectCountEstimate":33,"probeStatus":"ok","referenceUrl":"https://doi.org/10.3390/s21237853","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"josefheidler/har_adults_2021-harth\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"josefheidler/har_adults_2021-harth\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=josefheidler, license:mit, size_categories:1M<n<10M, format:parquet, downloads=430"],"format":"mixed","generatedAt":"2026-11-22T01:10:15.449Z","layout":{"example":"harmonized/agile-albatross.parquet","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"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","encoding":"hex","key":"harmonized/agile-albatross.parquet","truncatedAt":512}],"sampleObjects":["harmonized/agile-albatross.parquet","harmonized/amber-antelope.parquet","harmonized/blithe-boar.parquet","harmonized/breezy-bison.parquet","harmonized/crisp-crane.parquet"],"sidecars":[{"bytes":"---\npretty_name: \"HARTH — thigh + back accelerometry, healthy adults, activity recognition\"\nlicense: mit\nlanguage:\n  - en\ntags:\n  - accelerometry\n  - wearable\n  - human-activity-recognition\n  - time-series-classification\n  - thigh-worn\n  - physical-activity\n  - movement-behaviour\n  - time-series\ntask_categories:\n  - tabular-classification\n  - other\nsize_categories:\n  - 1M<n<10M\nconfigs:\n  - config_name: default\n    data_files:\n      - split: train\n        path: \"harmonized/*.parquet\"\n---\n\n# HARTH — thigh + back accelerometry, healthy adults, activity recognition\n\nDual-sensor accelerometry (Axivity AX3, 50 Hz, ±8 g) from the right thigh and lower\nback in 31 healthy adults, with activity ground truth for 12 activity types including\nwalking, running, stair climbing, cycling, and postural activities. Recordings are\n1–2.5 hours per participant of semi-free-living activity. Harmonized from the HARTH\nGitHub release into per-subject parquet — one row per raw accelerometer sample (50 Hz).\n31 participants, 8.3 M samples, >99.9 % labelled, 46.2 h labeled.\n\nDesigned to resemble the HUNT4 accelerometer data (Trøndelag Health Study, Norway,\n≈35,000 participants using the same dual-sensor setup at thigh and lower back), it\nserves as a labelled training resource for large-scale population health research.\n\n## Source\n\n- **Paper**: Logacjov, A., Bach, K., Kongsvold, A., Bårdstu, H. B., & Mork, P. J. (2021).\n  HARTH: A Human Activity Recognition Dataset for Machine Learning.\n  *Sensors*, *21*(23), 7853.\n  https://doi.org/10.3390/s21237853\n- **Related paper**: Bach, K., Kongsvold, A., Bårdstu, H., Bardal, E. M., Kjærnli, H. S., Herland, S., Logacjov, A., & Mork, P. J. (2022). A Machine Learning Classifier for Detection of Physical Activity Types and Postures During Free-Living. *Journal for the Measurement of Physical Behaviour*, *5*(1), 24–31. https://doi.org/10.1123/jmpb.2021-0015\n- **Raw data**: https://github.com/ntnu-ai-lab/harth-ml-experiments/tree/main/harth\n  (original UCI release: https://archive.ics.uci.edu/dataset/779/harth — 22 participants;\n  the GitHub version used here includes 31 participants with corrected labels)\n- **License**: MIT\n\n## Protocol\n\n- **Participants**: 31 healthy adults (8 female, 14 male in original 22-subject release;\n  age 38.6 ± 14 y, height 177.3 ± 8.3 cm, weight 72.9 ± 10.6 kg, BMI 23.1 ± 2.3)\n  (range: age 25–68 y, weight 56–92 kg, height 157–191 cm, BMI 19.2–28.4 kg/m²).\n- **Sensors**: Two Axivity AX3 per participant — right thigh (≈10 cm above the upper\n  border of the patella) and lower back (3rd lumbar vertebra), both mounted with USB\n  connector pointing down.\n  Attachment: a 5 × 7 cm moisture-permeable film (Opsite Flexifix, Smith & Nephew) on\n  the skin, sensor fixed with double-sided tape, covered by a second 10 × 8 cm film layer.\n- **Sample rate**: 50 Hz (100 Hz raw, published at 50 Hz).\n- **Recording — two sessions**:\n  - *Session 1* (22 participants, 6 female reported in protocol description): 1.5–2 h at the participant's workplace or\n    home (semi-free-living). Participants were asked to accumulate at least 2–3 min each\n    of sitting, standing, lying, walking, and running/jogging; no enforced order.\n    Total: ≈1,804 min (≈30 h) across 22 participants (average 120 ± 21.6 min/participant).\n  - *Session 2* (9 additional participants): 60 ± 9 min per participant (417.6 min /\n    ≈7 h total), focused on walking, running, and cycling across flat, uphill, and\n    downhill sections; incidental postures and stair-climbing also annotated.\n  - Raw data sampled at 100 Hz; downsampled to 50 Hz for the public release.\n\n### Ground-truth provenance\n\nFrame-by-frame video annotation using ANVIL software. Ground truth recorded with a\nGoPro Hero 3+ chest-mounted camera (30 fps, 640 × 360 pixels) pointing toward the\nfeet. Two independent human experts annotated each file; Fleiss' κ = 0.96. Sensor and\nvideo signals were aligned using three heel drops visible in both streams.\n\nPublished baseline (Logacjov et al. 2021, 12 labels, leave-one-subject-out):\nSVM F","encoding":"utf8","key":"README.md","name":"README.md","size":11196,"truncated":true}],"sourceObjectKey":"harmonized/agile-albatross.parquet","starterCell":"from datasets import load_dataset\nds = load_dataset(\"josefheidler/har_adults_2021-harth\")\nprint(ds)","starterSamplePath":"harmonized/agile-albatross.parquet","totalObjects":33},"summary":"HARTH, the Human Activity Recognition Trondheim dataset from the Norwegian University of Science and Technology (NTNU), in its 31-participant v2.0 version: raw triaxial…","title":"HARTH (Human Activity Recognition Trondheim) v2.0: Thigh and Lower-Back Accelerometry of 31 Adults with Video-Annotated Activities (NTNU)","updatedAt":"2026-11-22T04:04:56.763Z","visibility":"public","websiteUrl":"https://github.com/ntnu-ai-lab/harth-ml-experiments"}}