Human Activity Recognition Dataset for Pedestrians with Mobility Disabilities

Human Activity Recognition Dataset for Pedestrians with Mobility Disabilities

Note: The dataset is licensed and shared under Creative Commons Attribution 4.0 International (CC BY 4.0). To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/. If you want to refer to this dataset in a publication, please use the reference below: Author names: removed for anonymous review. 2024. Human Activity Recognition Dataset for Pedestrians with Mobility Disabilities Sci. Data 00, 000.

1     HAR-PMD Dataset

HAR-PMD(Human Activity Recognition for Pedestrians with Mobility Disabilities) dataset consists of smartphone and smartwatch sensor data from six pedestrian activities for people with mobility difficulties: still, walking, crutches, walker, manual wheelchairs, and electric wheelchairs. Sixty participants collected smartphone data, and sixty additional participants collected both smartphone and smartwatch data. Each activity was conducted in both indoor and outdoor environments. Thirteen smartphone sensors and five smartwatch sensors were collected. As a result, the dataset consists of 14,400 minutes of data from 120 participants.

For detailed descriptions of activities, environments, sensors, positions, data collection application, and data collection procedure, see METHODS in the paper.

2     DATA FORMAT

2.1   HAR-PMD Dataset

The dataset consists of 120 folders for each individual. Each folder contains metadata, smartphone, and smartwatch data if collected. Smartphone and smartwatch data are stored in comma-separated values (CSV) format files.

2.1.1 Metadata

• <user_id>/device.txt – smartphone model of the participant
• <user_id>/os_version.txt – Android OS version of the smartphone

2.1.2 Smartphone

• <user_id>/<device>/<user_id>_<activity>_<device>_<environment>.csv – collected smartphone sensor data
  • <user_id>: 1 – 120
  • <device>: phone
  • <activity>: still, walking, crutches, walker, manual (manual wheelchair), or electric (electric wheelchair)
  • <environment>: indoor or outdoor

2.1.3 Smartwatch

• <user_id>/<device>/<user_id>_<activity>_<device>_<environment>.csv – collected smartwatch sensor data
  • <user_id>: 61 – 120
  • <device>: watch
  • <activity>: still, walking, crutches, walker, manual (manual wheelchair), or electric (electric wheelchair)
  • <environment>: indoor or outdoor

2.2   Data Files

2.2.1 Smartphone sensor data

Table 1 describes each column in a smartphone sensor data CSV file. For more detailed descriptions, see Android Developers Sensors Documents.

<Table 1: Smartphone sensor data description for each column>

2.2.2 Smartwatch sensor data

Table 2 describes each column in a smartwatch sensor data CSV file. For more detailed descriptions, see Android Developers Sensors Documents.

<Table 2: Smartwatch sensor data description for each column>

3. Benchmark Experiment

We applied four machine learning models (Decision Tree, Random Forest, XGBoost, and Support Vector Machine) and four deep learning models (Multilayer Perceptron, Convolutional Neural Networks, Long Short-Term Memory (LSTM), and Transformer) to conduct the benchmark experiments.

We had three validation scenarios: a) Mobility disability; b) Walking aids & wheelchairs; and c) Mobility in detail. Data from all participants and activities were used for each validation scenarios. Each scenario was evaluated using two validation methods: user-depedent (random) evaluation and user-independent evaluation. In the user-dependent (random) evaluation, we randomly shuffled the data and divided the 14,400 min of data into five subsets of 2,880 min each for smartphone sensor data. After shuffling and splitting the data, one subset was allocated for testing while the other four subsets were used for training the model, and this process was repeated for all five subsets. For the combination of smartphone and smartwatch sensors, 7,200 min of data was randomly shuffled and divided the five subsets of 1,440 min data. Each subset was again allocated for testing while the other four subsets were used for training. In the user-independent (UI) evaluation, we used leave-one-group-out 5-fold cross-validation. For smartphone sensor data, 120 participants who comprised the data were divided into five equally distributed subsets of 24 participants each. The models were trained using the data of 96 participants and tested one of the data for the remaining 24 participants. For the combination of smartphone and smartwatch sensors, as the smartwatch data consisted of 60 participants, the data from 48 participants were used for training and the remaining 12 participants were used for testing. This list of participants included in each fold is shown in the below table. In the combination of sensor, we used three sensor combinations: a) linear accelerometer; b) linear accelerometer and gyroscope, and c) linear accelerometer, gyroscope, and magnetometer. Data from all participants and activities were used for each sensor scenarios.