|
|
|
| UvA | Fac. Science | IAS | Tim van Kasteren's Homepage | ||
|
Activities are a very important piece of information for ubiquitous applications. A sensor system capable of automatically recognizing activities would allow many potential applications in areas such as health care, comfort and security. For example, in elderly care, activities of daily living (ADLs) are used to assess the cognitive and physical capabilities of an elderly person. An activity recognition system allows us to automatically monitor their decline over time and detect anomalies. In this work we present a sensor network setup (fig. 1) and introduce an inexpensive and accurate method for annotation (software can be found below). We recorded a dataset consisting of 28 days of fully annotated sensor data (dataset can be found below). A number of experiments were conducted using this dataset, showing the effectiveness of Hidden Markov models (HMMs) and Conditional Random Fields (CRFs) in activity recognition. Experiments on the dataset were performed using a number of sensor data representations. Our raw sensor representation gives a 1 when the sensor is firing and a 0 otherwise (fig. 2a). Next to this raw data as observations we experiment with a change point representation. In this representation, the sensor gives a 1 to timeslices where the sensor reading changes (fig. 2b). Finally, we experiment with a last sensor fired representation, in which the last sensor that changed state continues to give 1 and changes to 0 when a different sensor changes state (fig. 2c). Results can be found in the UbiComp paper below. To download the data set, click here! |
|
