Extraction of transportation information from combined position and accelerometer tracks

Abstract

Travel surveys are increasingly taking advantage of global positioning system (GPS) data offering precise and objective route and time observations whilst potentially reducing response burden. However, there are still several open issues concerning the automated post-processing of these large datasets. Without a reliable post-processing, GPS-based studies require either a considerable amount of manual analysis, leading to costly surveys or extensive prompted-recall interviews with the respondents. As part of this thesis a travel diary study was conducted in the Greater Zurich Area. 150 participants carried dedicated GPS devices for up to one week and corrected their diaries in a web-based prompted recall tool. Using the resulting data set, the existing POSition DAta Processing framework was extended by a trip purpose module. Random forests, a machine learning technique, is used for classification. For trip purpose a share of correct predictions between 80 and 85 % is achieved for different setups. High variability in accuracy between persons is observed. Hence, personalisation strategies are tested. It is shown that the classifier is improved if it is learned on data that includes some of the participant?s annotations (median accuracy + 5.5 %). The updated processing tool, and also lessons learned from the GPS survey in Zurich are tested in the PEACOX project, a joint project with many partners where a smartphone cross modal trip planner was developed that encourages ecological friendly behaviour. GPS and accelerometer time series for 33 study participants in Vienna and Dublin are available for analysis; these were tracked simultaneously with smartphones and dedicated devices for 8 weeks. Therefore, further insight into the usefulness of smartphones and dedicated GPS devices for collecting current travel survey data is gained. Meaningful diaries can be extracted from both data sources. However, if high resolution data is needed, results suggest that dedicated GPS devices are still relevant; they have no battery issues, meaning that more data is recorded and that data quality is more stable. High resolution data is particularly interesting to observe taken routes. Two potential applications are shown here: route choice models are estimated for all travel modes (public transport, car, bicycle and walking) and parking search is shown to be hard to identify in our data.