De ontwikkeling van een akoestisch gebaseerd thuismonitor systeem ter verbetering van de thuiszorg

The pending retirement of the babyboom generation will put a strain on the elderly care support system. In particular, there will be a lack of qualified personnel to provide home care and also insufficient space in elderly care facilities. Technology integrated in the elderly’s home could help ease this burden. Firstly, this technology could keep an eye on things, reassuring the person's safety at home. In such a situation, the person can confidently live independently at home. This way the transfer to an elderly care facility can be postponed until a later stage in life. Secondly, the same technology could simultaneously gather various health-related information for the home caregivers to help diagnosing the elderly’s health situation. Furthermore, the daily gathering of this information makes it possible to track gradual changes over time which could otherwise be missed. This way the workload for both the home caregiver and the care facilities can be reduced. This topic will be addressed by developing and extending applications for a home monitoring system based on acoustic information gathered by microphones. Although many other sensors are also studied for home monitoring (camera, body-worn accelerometers, passive infrared sensor, etc.), acoustic information has some advantages over others: 1) Nowadays microphones are produced very cheaply. With prices for adequate microphones of a few euro per microphone, the total price for the home monitoring system can be limited. 2) The microphones do not have to be worn. This makes the system convenient for the user and limits its intrusion in his or her daily life. 3) Microphones do not depend on external lighting compared to cameras. Therefore the system’s performance will not suffer by bad lighting (at night, on a cloudy day, …). 4) Nowadays microphones are so small the system can be installed discretely in one’s home. 5) Some possible applications can only work on acoustic information, e.g. scream detection, assistive technology using speech recognition, etc. Furthermore, the system based on acoustic information can later be combined with e.g. cameras in a multi-sensor approach to achieve better performance. In this work a Wireless Acoustic Sensor Network (WASN) is used to gather the acoustic information. A WASN consists of multiple nodes equipped with one or more microphones and possibly a central processor. These nodes (and central processor) are wirelessly linked to exchange information. The use of multiple microphones has some advantages over a single microphone. Firstly, it allows the microphones to cover the whole room/house so that all possible acoustic sources have a microphone nearby for a qualitative observation of that source. Secondly, multiple recordings of the same acoustic source captured at different locations can be used to gather additional information about the source’s position and more noise reduction options are available. This doctoral research has made 2 main contributions to home monitoring systems based on acoustic information: (1) The design of a noise-robust footstep position detector. The detected footstep positions can be used to extract gait information which is known to be related to the person's health. For example, early stages of dementia or other cognitive impairment can be detected from the gait information. In this work standard sound source position estimation techniques are used, complimented by algorithms used to reduce the effect of the noise. It is shown that these complementary, noise-reducing, algorithms can reduce the absolute estimation error by 70%. In addition it will be shown that some noise-reducing algorithms are influenced by the type of noise. It is clear that some of the noise-reducing algorithms will fail under non-stationary noise but, by using a selection of these noise-reducing algorithms an estimation error reduction of 33% is still achieved. (2) Proposing a method to extract spatio-spectral features from multi-microphone data. These features can be used to extend existing activity detection applications. These applications try to detect current activities and are able to report abnormal, possibly alarming, situations. Typically activity detection applications use spectral features, describing how the observed signal sounds to a human listener. The proposed spatio-spectral features add an extra dimension to these, namely a spatial description of the sound (the origin of the observed sound). This way different activities can be better separated from their feature description. The proposed features are compared to the typical spectral features in 3 different experiments: a simulated sound source detection scenario, a real-life sound source detection scenario and a real-life activity detection scenario in the home of an older person. In almost all of the experiments the proposed features achieved a better detection score (up to 9.4% improvement), without ever achieving large disimprovements (at its worst 0.2% disimprovement).nrpages: 165status: publishe