Motion pattern analysis enabling accurate travel mode detection from GPS data only

Travel modes are one of the crucial pieces of information to characterize one´s travel behavior. In recent years several approaches of mode detection from GPS data have been proposed. The approach presented in this paper uses machine learning to evaluate a set of GPS-based features for their ability to recognize the common modes walk, bicycle, car, bus, and train. The proposed features describe motion characteristics from GPS-trajectories by relative frequencies. Compared to previous work the proposed feature set leads to higher average recognition rates around 92% without relying on additional GIS or real-time information. The evaluation compares detection rates from multilayer perceptrons, logistic model trees, and C4.5 decision trees and is complemented by an evolutionary feature selection for selecting the most beneficial feature subsets leading to the best quality gain. In contrast to other research, this study uses a comparatively large set of 400 GPS trajectories which have been recorded in rural and urban European areas. Results contribute to a higher reliability as well as a broader applicability of GPS-only travel mode detection.