To explore other possibilities, researchers at Intel use a small gadget, about the size of a pager, that amasses data from seven sensors: an accelerometer, a barometer, a humidity sensor, a thermometer, a light sensor, a digital compass, and a microphone, says Tanzeem Choudhury, a researcher at Intel Labs Seattle. Most of the sensors are used to determine location and activity, but the microphone can provide interesting insight into social networks, she says, such as whether a person is having a business conversation or a social chat. Aware of privacy concerns, the researchers designed the microphone data to be immediately processed so that all words are removed, and only information about tone, pitch, and volume is recorded. Recently, Intel researchers equipped a first-year class of University of Washington graduate students with these sorts of sensors and, based on their interactions, were able to watch social networks develop over time.
To churn through all the data the Intel sensors collect, the researchers designed software to process it in stages, explains Choudhury. "You can do some simple processing on the mobile device," she says, such as averaging similar data points over time and throwing out data from a sensor that's below a threshold. Most mobile phones have the processing capabilities to do this and extract actions such as walking and sitting.
In the next stage of processing, researchers plug these actions into machine-learning models that infer more-complex behaviors. For instance, making a meal will require short walking bursts, standing, and picking things up. The Intel researchers developed models that look for certain actions occurring in succession. These models can also adjust to the basic quirks of the user, accounting for variation in cooking behavior; some meals may require more walking than others, and some people may sit more during meal preparation than others. This sort of information could be useful,
Choudhury says, in determining if an elderly person is eating regularly. She notes that currently, some of the modeling is too computationally intensive to do entirely on a cell phone, and some of the data must be uploaded to a computer or a server. However, she says, the algorithms are becoming more efficient, and the processing power in phones continues to increase.
At this point, says MIT's Eagle, it wouldn't be too difficult to write consumer software that could infer a person's basic activities. These activities could then be used to update the status listed in an instant-messenger program or on a blog. Eagle notes, however, that manufacturers might be hesitant because it's likely that all the required data processing could cut battery life.
Apple has made no announcements about whether it might include such software in future versions of the iPhone. And it's unlikely that outside developers will be able to take advantage of the sensors at this point: Apple is limiting third-party development to applications that run within the Web browser--essentially, specialized Web pages. But as more phones become equipped with sensors, and phones' processing power continues to increase, Eagle suspects that sensor-based applications will become more popular.