A robust fall detection system is essential to support the independent living of elderlies. In this context, we develop a machine learning framework for fall detection and daily living activity recognition. Using acceleration data from public databases, we test the performance of two algorithms to classify seven different activities including falls and activities of daily living. We extract new features from the acceleration signal and demonstrate their effect on improving the accuracy and the precision of the classifier. Our analysis reveals that the quadratic support vector machine classifier achieves an overall accuracy of 93.2% and outperforms the artificial neural network algorithm.

This paper is concerned with the enhancement of the resolution of the spectrogram of non-stationary mobile radio channels using massive multiple-input multiple-output (MIMO) techniques. By starting from a new generic geometrical model for a non-stationary MIMO channel, we derive the complex MIMO channel gains under the assumption that the mobile station (MS) moves with time-variant speed. Closed-form solutions are derived for the spectrogram of the complex MIMO channel gains by using a Gaussian window. It is shown that the window spread can be optimized subject to the MS's speed change. Furthermore, it is shown that the spectrogram can be split into an auto-term and a cross-term. The auto-term contains the useful time-variant spectral information, while the cross-term can be identified as a sum of spectral interference components, which restrict considerably the time-frequency resolution of the spectrogram. Moreover, it is shown that the effect of the cross-term can be drastically reduced by using massive MIMO techniques. The proposed method is not only important for estimating timevariant Doppler power spectra with high resolution, but it also pioneers the development of new passive acceleration/deceleration estimation methods and the development of new non-wearable fall detection systems.

This paper proposes a three-dimensional (3D) non-stationary fixed-to-fixed indoor channel simulator model for human activity recognition. The channel model enables the formulation of temporal variations of the received signal caused by a moving human. The moving human is modelled by a cluster of synchronized moving scatterers. Each of the moving scatterers in a cluster is described by a 3D deterministic trajectory model representing the motion of specific body parts of a person, such as wrists, ankles, head, and waist. We derive the time-variant (TV) Doppler frequencies caused by the motion of each moving scatterer by using the TV angles of motion, angles of arrival, angles of departure. Moreover, we derive the complex channel gain of the received signal. Furthermore, we analyze the TV Doppler power spectral density of the complex channel gain by using the concept of the spectrogram and present its expression in approximated form. Also, we derive the TV mean Doppler shift and TV Doppler spread from the approximated spectrogram. The accuracy of the results is validated by simulations. The channel simulator is beneficial for the development of activity recognition systems with non-wearable devices as the demand for such systems has increased recently.

No abstract

No abstract