Specifically, our mode detection algorithm fuses inputs from the
mobile devices’ GPS receivers with real time locations of buses,
rail line and bus stop location data. GPS technology is a built-in
feature of many mobile devices, such as IPhones, BlackBerrys and
Android phones. Given a GPS trace of a traveler, one way to build
the classification model is as follows. For each GPS sensor report
in the trace, various features including the closest Euclidian
distance to rail lines, closest Euclidian distance to buses and closest
Euclidian distance to bus stops are computed. Mean speed,
heading, and acceleration are also obtained over a time window.
These features form a sensor feature vector. The feature vector,
plus the transportation mode label of the associated time interval, forms a training example. In this way, a training set is constructed.
This procedure is illustrated by Figure 2.