Frédéric Li¹, Kimiaki Shirahama¹, Muhammad Adeel Nisar¹, Lukas Köping¹, Marcin Grzegorzek^1,2

¹ Research Group for Pattern Recognition, University of Siegen, Germany
² Department of Knowledge Engineering, University of Economics in Katowice, Poland

Note (last updated on April 18, 2023): This webpage is a copy of the one linked in the paper, which was originally located on the website of Research Group for Pattern Recognition at the University of Siegen (Germany). Because of the movement of the second author who is managing this webpage, the old webpage at the web server of Kindai University (Japan) will no longer be accessible shortly. Consequently, all research materials used in the paper have been moved to the current institution of the second author Doshisha University (Japan). The contents contained and linked in this webpage are exactly the same as the original ones.
Note (last updated on May 10, 2019): This webpage is a copy of the one linked in the paper, which was originally located on the website of Research Group for Pattern Recognition at the University of Siegen (Germany). Because the whole research group moved to another University at the beginning of 2019, the old webpage will no longer be accessible in the coming months. Consequently, all research materials used in the paper have been moved to the web server of Kindai University (Japan), the current institution of the second author. The contents contained and linked in this webpage are exactly the same as the original ones.

The increasing availability of wearable sensors made the obtention of personal data easier, opening new opportunities of applications for Human Activity Recognition (HAR). The interest in this field is visible by the growing number of solutions - in particular deep-learning-based - proposed to learn effective feature representations for the classification of activities by analysing large amounts of data.

However, the solutions proposed in past research works can often prove to be hard to reproduce, due to unsufficient amounts of provided details regarding the data processing operations or setup of the feature learning models. In addition, the lack of a baseline evaluation framework makes a strict comparison between those techniques - as well as the establishment of a hierarchy among them - difficult.

To address those issues, an evaluation framework fixing the data pre-processing, segmentation and classifier construction phase is defined to test and compare the effectiveness of several state-of-the-art feature learning approaches including:

• Hand-crafted features
• Codebook-based approaches [code] (about 10MB)
• Multi-Layer-Perceptron (MLP)
• Long-Short-Term-Memory (LSTM)
• Hybrid CNN and LSTM models
• Autoencoders

The evaluation framework is used on two benchmarks for wearable-based HAR: the OPPORTUNITY and UniMiB-SHAR datasets. The results of the comparative study are summarized in our publication.

All research materials used for the studies, including raw and processed datasets, codes and instructions on how to use them, learned features, are provided on this webpage (see overview.txt). This includes:

• the data used in the study [link] (about 38GB)
• codes and instructions on how to use them [link] (about 10MB)
• the trained deep-learning models [link] (about 281MB)
• the features obtained on both datasets [link] (about 22GB)

For any question about the paper, study or code, please contact the main author of the publication, whose coordinates can be found here.

Please reference the following publication [1] if you use any of the codes and/or datasets linked above:

[1] Frédéric Li, Kimiaki Shirahama, Muhammad Adeel Nisar, Lukas Köping, Marcin Grzegorzek, Comparison of Feature Learning Methods for Human Activity Recognition using Wearable Sensors, Sensors, Vol. 18, No. 2, Article No. 679, 2018 (paper (open access))