Travel mode detection from GPS trajectories is actually
demanding for the researcher. The employed method should cope
with the complex relationship between the features that influence
the travel mode decision, but also take into consideration the
ambiguity that actually exists within these influential features.
More specifically, the most commonly used features, such as the
average speed, the average absolute acceleration and the travel
distance, are potentially interrelated. In addition, they present an
uncertainty in determining the travel modes. Furthermore, the
detection effort is particularly challenging when it is a necessary
part of the travel survey, as employed by most web-based GPS travel
surveys (Moiseeva, Jessurun, & Timmermans, 2010). In these
surveys, travel characteristics, including travel mode, are first
detected from GPS trajectories uploaded by respondents, and then
validated or corrected by them. Therefore, the detection accuracy
increases as new validated or corrected records are added to the
training set, implying that the survey system will learn over time
and thus reduce respondent burdens.