Abstract
In Human Activity Recognition (HAR), supervised
Machine Learning methods are predominantly used, making
availability of datasets a major issue for research in the field.
In particular, the majority of available datasets are collected
under controlled conditions. Consequently, models trained under
similar circumstances, generally exhibit a significant decrease in
recognition accuracy when they are moved to final deployment
in the wild, within unconstrained settings. This paper presents a
new dataset for HAR, collected in free-living and unconstrained
conditions. 10 subjects were recruited for the purpose of data
collection. Data was recorded over a 6 week period using
a smartphone app, and a wristband activity monitor. During
the first and last week of observation, participants also wore
an ActivPAL™ activity logger. The data collected have been
partially self-labeled by participants, by means of the mobile app
provided for data collection. The dataset collected can be used to
evaluate HAR algorithm and models in real-world unconstrained
settings. Together with the description of the dataset, this work
presents some preliminary results, obtained cross-validating a
model trained on the publicly available Extrasensory dataset, and
testing its performance on our newly collected dataset. Results
obtained highlighted high cross-subject variability when testing
on new subjects, with a balanced accuracy varying between
53.33% and 90.01%, with an average balanced accuracy of
71.73%.
Machine Learning methods are predominantly used, making
availability of datasets a major issue for research in the field.
In particular, the majority of available datasets are collected
under controlled conditions. Consequently, models trained under
similar circumstances, generally exhibit a significant decrease in
recognition accuracy when they are moved to final deployment
in the wild, within unconstrained settings. This paper presents a
new dataset for HAR, collected in free-living and unconstrained
conditions. 10 subjects were recruited for the purpose of data
collection. Data was recorded over a 6 week period using
a smartphone app, and a wristband activity monitor. During
the first and last week of observation, participants also wore
an ActivPAL™ activity logger. The data collected have been
partially self-labeled by participants, by means of the mobile app
provided for data collection. The dataset collected can be used to
evaluate HAR algorithm and models in real-world unconstrained
settings. Together with the description of the dataset, this work
presents some preliminary results, obtained cross-validating a
model trained on the publicly available Extrasensory dataset, and
testing its performance on our newly collected dataset. Results
obtained highlighted high cross-subject variability when testing
on new subjects, with a balanced accuracy varying between
53.33% and 90.01%, with an average balanced accuracy of
71.73%.
| Original language | English |
|---|---|
| Title of host publication | The 2nd Workshop on advanced Technologies for Smarter Assisted Living solutions: Towards an open Smart Home infrastructure (SmarterAAL'19) |
| Publisher | IEEE |
| Pages | 166-171 |
| Number of pages | 6 |
| Publication status | Accepted/In press - 21 May 2019 |
Keywords
- Balanced Batch Learning
- Convolutional Neural Networks
- Dataset
- Free-living
- Human Activity Recognition
- Machine Learning
