Travel mode estimation for multi-modal journey planner

• Published in: Transportation research. Part C, Emerging technologies Vol. 82; pp. 273 - 289
• Main Authors: Mäenpää, Heikki, Lobov, Andrei, Martinez Lastra, Jose L.
• Format: Journal Article
• Language: English
• Published: Elsevier India Pvt Ltd 01.09.2017
• Subjects: Bayes classifier; Feature selection; Neural network; Travel mode detection; Travel mode detection; Bayes classifier; Feature selection; Neural network
• ISSN: 0968-090X; 1879-2359
• DOI: 10.1016/j.trc.2017.06.021

•Travel mode recognition based on location data at 1 sample/min.•Sub-8mHz spectrum is useful for travel mode recognition.•Skewnesses and kurtoses of speed and acceleration are not useful.•Auto- and cross correlations at 1 sample/min are also not useful. For route planning and tracking, it is sometimes necessary to know if the user is walking or using some other mode of transport. In most cases, the GPS data can be acquired from the user device. It is possible to estimate user’s transportation mode based on a GPS trace at a sampling rate of once per minute. There has been little prior work on the selection of a set of features from a large number of proposed features, especially for sparse GPS data. This article considers characteristics of distribution, auto- and cross-correlations, and spectral features of speed and acceleration as possible features, and presents an approach to selecting the most significant, non-correlating features from among those. Both speed and acceleration are inferred from changes in location and time between data points. Using GPS traces of buses in the city of Tampere, and of walking, biking and driving from the OpenStreetMap and Microsoft GeoLife projects, spectral bins were found to be among the most significant non-correlating features for differentiating between walking, bicycle, bus and driving, and were used to train classifiers with a fair accuracy. Auto- and cross-correlations, kurtoses and skewnesses were found to be of no use in the classification task. Useful features were found to have a fairly large (>0.4) correlation with each other.