Affiliations: [a] Departmento de Electrónica y Tecnología de Computadoras, Universidad Politécnica de Cartagena, Cartagena, Spain | [b] Ambient Intelligence and Interaction, SLL, Campus Universitario de Espinardo, Murcia, Spain
Correspondence:
[*]
Corresponding author. Juan Zapata, Departmento de Electrónica y Tecnología de Computadoras, Universidad Politécnica de Cartagena, Antiguo Cuartel de Antigones. Campus Muralla del Mar. Plaza del Hospital 1, 30202 Cartagena, Spain. E-mail:[email protected]
Abstract: Occupancy detection in customary resting places of people living alone at home is a main issue in Ambient Assisted Living (AAL) systems.
Typical occupancy sensors are based on pressure-contact mats or load transducers with a predefined threshold which is established during the sensor installation and should be recalibrated manually if the transducer response changes. We propose an automatic calibration algorithm to get an adaptive threshold which makes the sensor robust regarding the transducer degradation, a small shift in the position of the sensor, or any other variation that modifies the characteristic response of the load sensor. An implementation of the k-means clustering algorithm is used over a data window to get adaptive centroids that represent the values associated with each sensor state: occupied and unoccupied. The load threshold for occupancy determination is updated using these centroids. This method has demonstrated to have better percentage of success in the classification between occupancy and no occupancy (success rate) and better error margin than the fixed one.
The improvement brought by the proposed adaptive method has been quantified using load data records from occupancy sensors in an AAL system deployed over 10 real homes.
Success rates have been satisfactorily increased in every case, as well as error margin, which has been measured by defining a metric for the mean of the deviation of the threshold from the center of the local data.
The success rate has been increased from 77.7% to 96% in the best case and from 93% to 98.6% in mean. The deviation of the threshold from the center of the local data, as an inverse measurement of the error margin, has been improved from 24.4% to 0.75% in the best case and from 9.5% to 1.14% in mean.
