Improving Activity Recognition Accuracy in Ambient-Assisted Living Systems by Automated Feature Engineering

• Autor(a) principal: Zdravevski, Eftim
• Data de Publicação: 2017
• Outros Autores: Lameski, Petre, Trajkovik, Vladimir, Kulakov, Andrea, Chorbev, Ivan, Goleva, Rossitza, Pombo, Nuno, Garcia, Nuno M.
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.6/8267
• topic: Feature extraction; Time series analysis; Ambient intelligence; Wearable sensors; Sensor fusion; Pattern recognition; Data mining; Data preprocessing; Body sensor networks

Ambient-assisted living (AAL) is promising to become a supplement of the current care models, providing enhanced living experience to people within context-aware homes and smart environments. Activity recognition based on sensory data in AAL systems is an important task because 1) it can be used for estimation of levels of physical activity, 2) it can lead to detecting changes of daily patterns that may indicate an emerging medical condition, or 3) it can be used for detection of accidents and emergencies. To be accepted, AAL systems must be affordable while providing reliable performance. These two factors hugely depend on optimizing the number of utilized sensors and extracting robust features from them. This paper proposes a generic feature engineering method for selecting robust features from a variety of sensors, which can be used for generating reliable classi cation models. From the originally recorded time series and some newly generated time series [i.e., magnitudes, rst derivatives, delta series, and fast Fourier transformation (FFT)-based series], a variety of time and frequency domain features are extracted. Then, using two-phase feature selection, the number of generated features is greatly reduced. Finally, different classi cation models are trained and evaluated on an independent test set. The proposed method was evaluated on ve publicly available data sets, and on all of them, it yielded better accuracy than when using hand-tailored features. The bene ts of the proposed systematic feature engineering method are quickly discovering good feature sets for any given task than manually nding ones suitable for a particular task, selecting a small feature set that outperforms manually determined features in both execution time and accuracy, and identi cation of relevant sensor types and body locations automatically. Ultimately, the proposed method could reduce the cost of AAL systems by facilitating execution of algorithms on devices with limited resources and by using as few sensors as possible.