ctrl-gaurav / harth-ml-experiments

Baseline Machine Learning models for the Human Activity Recognition Trondheim (HARTH) dataset

HAR Datasets and Machine Learning experiments

Baseline Machine Learning models for the Human Activity Recognition Trondheim (HARTH) and the Human Acceleration Recognition 70+ (HAR70+) datasets, proposed and used in our papers: HARTH: A Human Activity Recognition Dataset for Machine Learning, A Machine Learning Classifier for Detection of Physical Activity Types and Postures During Free-Living, and Validation of an Activity Type Recognition Model Classifying Daily Physical Behavior in Older Adults: The HAR70+ Model.

HARTH Dataset

The folder harth contains the Human Activity Recognition Trondheim Dataset (HARTH). It consists of acceleration data of 22 subjects, which wore two three-axial Axivity AX3 (Axivity Ltd., Newcastle, UK) accelerometers on the thigh and lower back. The dataset is also uploaded to the UC Irvine Machine Learning Repository.

Setup

• Acceleration signals
• 2 three-axial Axivity AX3 accelerometers
• Attached to: thigh and lower back

Activity Annotations

Label	Activity	Notes
1	walking
2	running
3	shuffling	standing with leg movement
4	stairs (ascending)
5	stairs (descending)
6	standing
7	sitting
8	lying
13	cycling (sit)
14	cycling (stand)
130	cycling (sit, inactive)	cycling (sit) without leg movement
140	cycling (stand, inactive)	cycling (stand) without leg movement

HAR70+ Dataset

The folder har70plus contains the Human Activity Recognition 70+ (HAR70+) dataset. It consists of acceleration data of 18 older-adult subjects, which wore two three-axial Axivity AX3 (Axivity Ltd., Newcastle, UK) accelerometers on the thigh and lower back. The dataset is also uploaded to the UC Irvine Machine Learning Repository.

Setup

• Acceleration signals
• 2 three-axial Axivity AX3 accelerometers
• Attached to: thigh and lower back

Activity Annotations

Label	Activity	Notes
1	walking
3	shuffling	standing with leg movement
4	stairs (ascending)
5	stairs (descending)
6	standing
7	sitting
8	lying

Machine Learning Experiments

The folder experiments contains all our experiments. It is possible to train a K-Nearest Neighbors, a Support Vector Machine, a Random Forest, an Extreme Gradient Boost, a Convolutional Neural Network, a Bidirectional Long Short-term Memory, and a CNN with multi-resolution blocks.

Requirements

• Python 3.8.10+

cd experiments
pip install -r requirements.txt

Usage

Start a model training using HARTH

cd experiments
./run_training.sh -c <path/to/model/config.yml> -d <path/to/dataset>
# Example: ./run_training.sh -c traditional_machine_learning/params/xgb_50hz/config.yml -d ../harth/

Each model can be configured using the corresponding config.yml file: xgb, svm, rf, knn, cnn, multi_resolution_cnn, lstm

Citation

If you use the HARTH dataset for your research, please cite the following papers:

@article{logacjovHARTHHumanActivity2021,
  title = {{{HARTH}}: {{A Human Activity Recognition Dataset}} for {{Machine Learning}}},
  shorttitle = {{{HARTH}}},
  author = {Logacjov, Aleksej and Bach, Kerstin and Kongsvold, Atle and B{\aa}rdstu, Hilde Bremseth and Mork, Paul Jarle},
  year = {2021},
  month = nov,
  journal = {Sensors},
  volume = {21},
  number = {23},
  pages = {7853},
  publisher = {{Multidisciplinary Digital Publishing Institute}},
  doi = {10.3390/s21237853}
}

@article{bachMachineLearningClassifier2021,
  title = {A {{Machine Learning Classifier}} for {{Detection}} of {{Physical Activity Types}} and {{Postures During Free-Living}}},
  author = {Bach, Kerstin and Kongsvold, Atle and B{\aa}rdstu, Hilde and Bardal, Ellen Marie and Kj{\ae}rnli, H{\aa}kon S. and Herland, Sverre and Logacjov, Aleksej and Mork, Paul Jarle},
  year = {2021},
  month = dec,
  journal = {Journal for the Measurement of Physical Behaviour},
  pages = {1--8},
  publisher = {{Human Kinetics}},
  doi = {10.1123/jmpb.2021-0015},
}

If you use the HAR70+ dataset for your research, please cite the following paper:

@article{ustadValidationActivityType2023,
  title = {Validation of an {{Activity Type Recognition Model Classifying Daily Physical Behavior}} in {{Older Adults}}: {{The HAR70}}+ {{Model}}},
  shorttitle = {Validation of an {{Activity Type Recognition Model Classifying Daily Physical Behavior}} in {{Older Adults}}},
  author = {Ustad, Astrid and Logacjov, Aleksej and Trolleb{\o}, Stine {\O}verengen and Thingstad, Pernille and Vereijken, Beatrix and Bach, Kerstin and Maroni, Nina Skj{\ae}ret},
  year = {2023},
  month = jan,
  journal = {Sensors},
  volume = {23},
  number = {5},
  pages = {2368},
  publisher = {{Multidisciplinary Digital Publishing Institute}},
  issn = {1424-8220},
  doi = {10.3390/s23052368},
  copyright = {http://creativecommons.org/licenses/by/3.0/}
}

Note

Our HARTH dataset is subject to changes in future releases. Therefore, consider version v1.0 for reproducibility purposes. It contains the dataset and experiments used in our article, HARTH: A Human Activity Recognition Dataset for Machine Learning