ActiMotus Validation — Reproducibility Package

Reproduces the validation of the ActiMotus human activity recognition algorithm against video ground truth, on five public accelerometry datasets covering children, adults and older adults.

Three commands regenerate every table and confusion matrix in the study. No data is redistributed here: it is fetched from HuggingFace at pinned revisions and the algorithm comes from PyPI, so the package supplies only the pipeline between them. The generated tables and figures are committed under results/ so they can be read without running anything; re-running overwrites them.

Run it

uv sync
uv run python scripts/01_features.py    # ~2 GB download, tens of minutes
uv run python scripts/02_activities.py  # seconds
uv run python scripts/03_analysis.py    # seconds

Results land in results/ as .xlsx tables and .png confusion matrices. Add --dataset <name> to restrict stages 1–2 to one dataset, --limit N for a quick smoke run, --only <stem> to rebuild a single stage 3 output.

Each stage stamps its output with the ActiMotus version and dataset revisions it used, and refuses to consume a cache built from different inputs — a stale cache is an error, not a source of plausible wrong numbers.

Datasets

Dataset	Population	Sensors	HuggingFace	License
Lendt Adults	35 adults	lateral thigh, SENS 12.5 Hz	josefheidler/har_adults_2024-lendt	CC-BY-4.0
NTNU Children	46 typically-developing children	thigh + back, AX3 50 Hz	josefheidler/har_children_2024-harth	CC0-1.0
NTNU Adults	31 adults	thigh + back, AX3 50 Hz	josefheidler/har_adults_2021-harth	MIT
NTNU Older Adults	18 adults aged 70–95	thigh + back, AX3 50 Hz	josefheidler/har_older-adults_2023-harth	CC-BY-4.0
NTNU Walking Speeds	24 adults	thigh + back, AX3 50 Hz	josefheidler/har_ws_adults_2025-harth	CC-BY-4.0

Revisions are pinned in datasets.toml. This package redistributes no data; the dataset licenses bind you at download, and CC-BY-4.0 requires attribution to the source study listed on each dataset card.

Results

F1 per activity against video ground truth, ActiMotus 2.3.3 with its built-in DEFAULT thresholds, thigh sensor only:

Dataset	Lie	Sit	Stand	Shuffle	Walk	Stairs	Run	Cycle
Lendt, laboratory	1.00	1.00	0.99	—	0.99	—	1.00	1.00
Lendt, free-living	0.08	0.90	0.77	0.43	0.89	0.35	0.97	0.97
NTNU Children	0.79	0.83	0.81	0.33	0.88	0.45	0.82	0.88
NTNU Adults	0.08	0.79	0.79	0.38	0.84	0.62	0.93	0.89
NTNU Older Adults	0.05	0.82	0.83	0.29	0.86	0.18	—	—
NTNU Walking Speeds	—	—	—	—	0.79	—	0.98	—

Walking speeds is the only protocol that separates fast walking from walking, since speed cannot be established from free-living video: F1 0.79 for walking and 0.65 for fast walking. Elsewhere the two are pooled as walk.

Rows are normalised over the true class. The remaining figures — fused classes, the thigh + trunk configuration and walking speeds — are in results/, alongside the .xlsx tables and a provenance.json recording the ActiMotus version and dataset revisions that produced them.

Adding the lower-back sensor changes only lying and sitting; every other activity is unchanged to two decimals, since the trunk feeds only that discrimination:

Dataset	Lie, thigh	Lie, +trunk	Sit, thigh	Sit, +trunk
NTNU Children	0.79	0.90	0.83	0.93
NTNU Adults	0.08	0.90	0.79	0.91
NTNU Older Adults	0.05	0.77	0.82	0.87

Fused classes

Collapsing to five behaviour classes — sedentary (lying + sitting), standing (standing + shuffling), walking (walking + fast walking + stairs), running and cycling — gives F1, thigh sensor only:

Dataset	Sedentary	Standing	Walking	Running	Cycling
Lendt, laboratory	1.00	0.99	0.99	1.00	1.00
Lendt, free-living	0.99	0.83	0.91	0.97	0.97
NTNU Children	0.97	0.85	0.89	0.82	0.88
NTNU Adults	0.95	0.82	0.85	0.93	0.89
NTNU Older Adults	0.99	0.84	0.91	—	—

The lower-back sensor makes no difference here — every fused F1 is unchanged to two decimals with or without it. Its whole contribution is separating lying from sitting, and both collapse into sedentary.

Precision, recall and F1 with 90% confidence intervals, computed per participant and then averaged across participants, are written to the .xlsx tables in results/, alongside confusion matrices as .png, for both the eight-activity and fused vocabularies.

Sensor orientation

The published datasets use a hub frame with x up along the limb, y right and z forward. The two sensors need opposite treatment, so data.sensor_frame takes a required to_acti_frame argument rather than guessing:

• Thigh — rotated 180° about its long axis (y and z negated). ActiMotus expects the thigh z posterior. Without this, Lendt falls from 0.952 to 0.665.
• Back — left as published. ActiMotus expects the trunk z anterior, which the hub frame already provides.

Automatic flip detection stays enabled, but only as a guard against genuinely mis-worn sensors — never as a substitute for the conversion. With the frames correct it changes nothing on 153 of 154 thigh recordings, and both orientation=True and orientation=False give identical predictions, an invariant enforced by tests/test_integration.py.

Known limitations

• Thresholds are ActiMotus's built-in DEFAULT configuration, tuned by Bayesian optimization outside this package. Re-deriving them is out of scope.
• Lying detection from the thigh alone fails for adults and older adults (recall 0.06 in both); the back sensor resolves it (0.87 and 0.78). Children are the exception, reaching 0.97 from the thigh alone.

Citing

Please cite the validation study, not this repository. The article is under review; its DOI will be added here on publication.

If you need to reference the code specifically — for example to pin the exact version that produced a result — this package also has a DOI, 10.5281/zenodo.21955041, which always resolves to the latest release.

The datasets are cited separately. Each HuggingFace card names the study to attribute; CC-BY-4.0 requires it.

License

BSD 3-Clause, matching ActiMotus. See LICENSE.