Posture Recognition for Fall Prevention with Low-Complexity UWB WBANs

Abstract

This thesis concerns wearable ultra-wideband (UWB) wireless systems for human posture recognition. Our main objective is a technical proposal for a low-complexity realization of such a system. Posture recognition is a key enabler of various applications in the emerging field of technology-assisted approaches to maintain or improve human health. Especially fall prevention in the context of assisted living for elderly citizens relies on accurate body posture monitoring to detect and prevent imminent falls. Wireless body area networks (WBANs) pose a suitable means for continuous posture monitoring while preserving the user’s privacy. Such wireless systems directly offer secondary uses such as sensor data transmission. The difficulty of accurately modeling the time-varying wireless channels around the human body hinders analytical solution attempts to posture recognition. This thesis follows a measurement-based approach to posture recognition. For this purpose, we conduct a comprehensive measurement campaign. Therein we acquire the channel matrix between 18 body-mounted wireless nodes for a wide frequency range and a wide variety of postures. The postures are selected to cover an extensive range of daily activities. Measurements are performed in different environments for various test subjects of different physique in order to ensure a diverse and representative dataset. The acquired measurement data provides a solid foundation for a feasibility analysis and selection of key parameters of a posture recognition system for fall prevention from WBAN signals. The constraints of a wearable battery-powered system require the on-body wireless hardware to be of low complexity. Based on the acquired high-resolution measurement data, we distinguish three different levels of measured raw data corresponding to different hardware complexity levels. In particular, we consider the complex-valued channel impulse responses, their magnitude, and the received signal energy for each on-body link. We evaluate suitable physical layer options to measure the respective information for each level, which is subsequently used for posture recognition. In the course of a preliminary feasibility analysis, we develop a simplified system model and provide the corresponding maximum likelihood classifiers for the posture recognition problem for all three levels of measured data. Our analysis shows that posture classification based on the acquired data is feasible. It furthermore allows us to define an operating range of crucial parameters for the further system design. For the classification of postures based on channel measurements, we consider a variety of standard machine learning algorithms. Based on their quantitative comparison, Random Forests are selected as a suitable classification method which can identify postures reliably from low-complexity energy measurements. In order to further reduce the complexity and hardware demands, we minimize the number of required WBAN nodes. We demonstrate that accurate posture classification is possible with few suitably placed low-complexity nodes. This low-cost and low-complexity system design is thoroughly tested regarding its robustness towards additional disturbances and limitations of the data. The analysis includes an evaluation of the influence of the environment, cross-subject testing, and the reliability for the classification of unknown postures which are not present in the training data. Based on the findings of this work, we propose a conceptual design for a wearable posture recognition system. We outline the requirements and limitations of this design proposal for fall prevention applications and provide a realistic performance assessment of its implementation.