WHAR Datasets

Overview

This library offers comprehensive support for widely used WHAR (Wearable Human Activity Recognition) datasets, including:

• automated downloading from original sources and data extraction
• parsing into a unified, standardized data format
• configurable pre-processing (e.g., resampling, windowing) and post-processing (e.g., normalization)
• dataset splitting for common evaluation protocols such as LOSO and K-Fold cross-validation
• built-in caching and multi-processing for improved efficiency
• seamless integration with PyTorch and TensorFlow

The library currently includes out-of-the-box support for 37 datasets (listed below). Additional WHAR datasets can be easily integrated by defining a custom configuration with an associated parser and registering it with the framework.

Notice

This library does not host any datasets. To use a dataset, please visit its original website and make sure you understand and agree to the dataset’s terms and conditions.

Installation

With uv:

uv add "whar-datasets @ git+https://github.com/teco-kit/whar-datasets.git"

With optional TensorFlow support:

uv add "whar-datasets[tf] @ git+https://github.com/teco-kit/whar-datasets.git"  

With pip:

pip install "whar-datasets @ git+https://github.com/teco-kit/whar-datasets.git"

With optional TensorFlow support:

pip install "whar-datasets[tf] @ git+https://github.com/teco-kit/whar-datasets.git"

Quickstart

from whar_datasets import (
    Loader,
    LOSOSplitter,
    PostProcessingPipeline,
    PreProcessingPipeline,
    TorchAdapter,
    WHARDatasetID,
    get_dataset_cfg,
)

# create cfg for WISDM dataset
cfg = get_dataset_cfg(WHARDatasetID.WISDM)

# create and run pre-processing pipeline
pre_pipeline = PreProcessingPipeline(cfg)
activity_df, session_df, window_df = pre_pipeline.run()

# create LOSO splits
splitter = LOSOSplitter(cfg)
splits = splitter.get_splits(session_df, window_df)
split = splits[0]

# create and run post-processing pipeline for the specific split
post_pipeline = PostProcessingPipeline(cfg, pre_pipeline, window_df, split.train_indices)
samples = post_pipeline.run()

# create dataloaders for the specific split
loader = Loader(activity_df, session_df, window_df, post_pipeline.samples_dir, samples)
adapter = TorchAdapter(cfg, loader, split)
dataloaders = adapter.get_dataloaders(batch_size=64)

Supported Datasets

Single-Sensor Datasets

Supported	Name	Year	Paper	Citations
✅	WISDM	2010	Activity Recognition using Cell Phone Accelerometers	3862
✅	UCI-HAR	2013	A Public Domain Dataset for Human Activity Recognition using Smartphones	3372
✅	UTD-MHAD	2015	UTD-MHAD: A Multimodal Dataset for Human Action Recognition Utilizing a Depth Camera and a Wearable Inertial Sensor	997
✅	HAPT	2016	Transition-aware human activity recognition using smartphones.	939
✅	USC-HAD	2012	USC-HAD: A Daily Activity Dataset for Ubiquitous Activity Recognition Using Wearable Sensors	753
✅	UniMiB-SHAR	2017	Unimib shar: a dataset for human activity recognition using acceleration data from smartphones	712
✅	MotionSense	2019	Mobile Sensor Data Anonymization	345
✅	RealLifeHAR	2020	A Public Domain Dataset for Real-Life Human Activity Recognition Using Smartphone Sensors	208
✅	WISDM-19-PHONE	2019	WISDM: Smartphone and Smartwatch Activity and Biometrics Dataset	198
✅	WISDM-19-WATCH	2019	WISDM: Smartphone and Smartwatch Activity and Biometrics Dataset	198
✅	KU-HAR	2021	KU-HAR: An open dataset for heterogeneous human activity recognition	187
✅	Hang-Time	2023	Hang-time HAR: A benchmark dataset for basketball activity recognition using wrist-worn inertial sensors	52
✅	CAPTURE-24	2024	CAPTURE-24: A large dataset of wrist-worn activity tracker data collected in the wild for human activity recognition	45
✅	HARSense	2021	Harsense: statistical human activity recognition dataset	5
✅	FallDet	-	-	-

Multi-Sensor Datasets

Supported	Name	Year	Paper	Citations
✅	ActRecTut-Gestures	2014	A tutorial on human activity recognition using body-worn inertial sensors.	2086
✅	ActRecTut-Walking	2014	A tutorial on human activity recognition using body-worn inertial sensors.	2086
✅	PAMAP2	2012	Introducing a New Benchmarked Dataset for Activity Monitoring	1758
✅	OPPORTUNITY	2010	Collecting complex activity datasets in highly rich networked sensor environments	1024
✅	HHAR	2015	Smart Devices are Different: Assessing and Mitigating Mobile Sensing Heterogeneities for Activity Recognition	1019
✅	MHEALTH	2014	mHealthDroid: A Novel Framework for Agile Development of Mobile Health Applications	887
✅	DSADS	2010	Comparative study on classifying human activities with miniature inertial and magnetic sensors	780
✅	SAD	2014	Fusion of Smartphone Motion Sensors for Physical Activity Recognition	752
✅	BMHAD	2013	Berkeley MHAD: A Comprehensive Multimodal Human Action Database	668
✅	Daphnet	2009	Ambulatory monitoring of freezing of gait in Parkinson’s disease	652
✅	SKODA	2008	Wearable activity tracking in car manufacturing	504
✅	RealWorld	2016	On-body Localization of Wearable Devices: An Investigation of Position-Aware Activity Recognition	482
✅	UP-Fall	2019	UP-fall detection dataset: A multimodal approach	462
✅	UMAFall	2017	Umafall: A multisensor dataset for the research on automatic fall detection	243
✅	REALDISP	2014	Dealing with the Effects of Sensor Displacement in Wearable Activity Recognition	216
✅	HuGaDB	2018	HuGaDB: Human Gait Database for Activity Recognition from Wearable Inertial Sensor Networks	154
✅	HARTH	2021	HARTH: A Human Activity Recognition Dataset for Machine Learning	132
✅	w-HAR	2020	w-HAR: An Activity Recognition Dataset and Framework Using Low-Power Wearable Devices	100
✅	WEAR	2024	Wear: An outdoor sports dataset for wearable and egocentric activity recognition	66
✅	HAR70+	2021	A machine learning classifier for detection of physical activity types and postures during free-living	55
✅	UCA-EHAR	2022	UCA-EHAR: A Dataset for Human Activity Recognition with Embedded AI on Smart Glasses	35
✅	GOTOV	2022	A recurrent neural network architecture to model physical activity energy expenditure in older people	33

Configuration Reference

The easiest way to start is to load a built-in configuration and adjust the fields you need:

from whar_datasets import WHARDatasetID, get_dataset_cfg

cfg = get_dataset_cfg(WHARDatasetID.WISDM, datasets_dir="./datasets/")
cfg.parallelize = True
cfg.window_time = 3.0
cfg.window_overlap = 0.5
...

get_dataset_cfg(...) returns a dataset-specific WHARConfig with sensible defaults. The most useful fields are:

Field	Meaning	Default
datasets_dir	Root directory used to cache downloads, extracted files, metadata, windows, and samples.	./datasets/
in_memory	Whether post-processing keeps samples in memory or loads them from disk when needed.	True
parallelize	Enables parallel preprocessing steps.	False
cache_each_split	Caches split-specific samples separately so repeated runs can reuse them.	True
selected_activities	Optional activity filter applied before windowing.	None
selected_channels	Optional channel filter applied before windowing.	None
window_time	Sliding window length in seconds.	3.0
window_overlap	Window overlap ratio.	0.5
resampling_freq	Optional resampling rate in Hz before windowing.	None
val_percentage	Fraction of training data reserved for validation.	0.2
num_subject_groups	Number of groups used for leave-group-out splitting.	10
num_folds	Number of folds used for K-fold splitting.	10
normalization	Normalization strategy used in post-processing.	STD_GLOBALLY
transform	Optional transform applied to windows, such as STFT or DWT.	None
batch_size	Default batch size used by adapters and training helpers.	64
learning_rate	Default learning rate for downstream training.	1e-4
num_epochs	Default number of training epochs.	100
seed	Random seed used for sampling and dataloader shuffling.	0

If you want to benchmark multiple built-in datasets, BENCHMARK_DATASET_IDS contains the curated subset used by the library.

Citation

If you use the WHAR Datasets library in your research, please cite our paper:

@inproceedings{burzer2025whar,
  title={WHAR Datasets: An Open Source Library for Wearable Human Activity Recognition},
  author={Burzer, Maximilian and King, Tobias and Riedel, Till and Beigl, Michael and R{\"o}ddiger, Tobias},
  booktitle={Companion of the 2025 ACM International Joint Conference on Pervasive and Ubiquitous Computing},
  pages={1315--1322},
  year={2025}
}