The accuracy of GPS varies. For example, GPS tends to
underperform if it does not have a clear view of the sky (e.g. in
urban canyons). For this reason, we perform a noise filtering step
before training the classifier. Invalid GPS points are suppressed
based on the GPS accuracy and the change in speed. GPS sensor
reports with high inaccuracy readings and unrealistic changes in
speed are pruned. This is a manual step before classifier training.
GPS noise filtering before classifier training is not a new concept.
The authors of [4] perform a preprocessing step before training
their classifier.