Most approaches to recognising activities such as walking and traveling in a car use an accelerometer. Whilst this has proved successful it requires the additional accelerometer hardware, something not present on today’s mobile phones. We have therefore developed an alternative method of inferring activity using GSM data present on all GSM mobile phones. We utilise the fact that the patterns of cell and signal strength fluctuation behave in slightly different ways depending upon the current activity of the carrier of the mobile phone. For example, in Figure 1 you can see the sum of signal strength fluctuation across all visible cells. That is, the total amount of signal strength variance observed over 15-second time intervals. This figure shows it is relatively easy to distinguish between walking and remaining stationary by looking at signal strength fluctuation alone. However, at times, walking and traveling in a motor car share similar signal strength patterns. By comparing the fluctuation with the GPS traces we found that the drops between high spikes of fluctuation typically occurred whilst waiting at areas of traffic flow control or road junctions. Hence the graph reflects the stop-start nature of driving in metropolitan environments.  The GSM data provides us with an indication of the activity, but this needs to be smoothed out by knowledge of “normal” behaviour. For example, it is usual for a person to drive for a prolonged period of time and then to walk; it is unusual for a person to frequently switch between driving and sitting. We can model this using a Hidden Markov Model. In Table 1 we present results using data collected from a metropolitan environment. This confusion matrix shows that sensing walking and remaining stationary is relatively easy but sensing driving is much harder. This is due to the stop-start nature of driving in metropolitan environments.
Whilst not as fast and accurate as an accelerometer we believe that this low cost approach to sensing activity will support a multitude of pervasive applications. The primary advantage this approach offers over an accelerometer based method to sense activity is that it does not require any additional sensor hardware. As a proof of concept we have implemented this work on Orange SPV C500 mobile phones. We have found that both the calibration and running of the HMM are able to occur in real time. |