Detection and classification of acoustic events for in-home care

Due to the demographic change, the number of older people will grow steadily within the next decades [1]. Technical systems, methods and algorithms developed in the context of Ambient Assisted Living (AAL) aim to enable these people to live self-determined and safe in their own homes as long as possible. To ensure this, these homes may have to be equipped with e.g., a health monitoring system, special sensors or devices for gathering and evaluating individual information. In this paper, an acoustic monitoring system for evaluation of acoustic cues is presented and tested exemplarily on four acoustic events (cough, knock, clap and phone bell). The acoustic monitoring includes several consecutive signal processing steps: the audio signal picked up by a microphone is transformed in a psycho-physiologically weighted domain using a gammatone filterbank, resulting in a so-called cochleogram [2]. From this cochleogram, background noise is estimated and eliminated. Further signal processing determines partitions of the cochleogram which are considered to form acoustic events. Each of these events is then evaluated with a classification algorithm. In this classification step, class membership is determined by comparing the event under inspection to representatives which have been calculated during a training phase utilizing a self-organizing map [3]. It is shown that a true positive rate from over 79% can be achieved where the false positive rate is smaller than 4% except for the event knocking.