Inferring Individual Trip Chains from Smartphone-based GPS Data

To address the limitations of current travel data, such as high-labor cost, inaccuracy in time and location, and missing trips, we develop a smartphone-based travel survey system and propose a method for inferring individual trip chains from smartphone GPS data with all-period, multimodal, and complete trips. Firstly, anchors are extracted from personal trajectories with a proposed spatiotemporal density-based clustering algorithm. The anchors are then classified into public transport transfer nodes and trip ends using a random forest model. Secondly, a spatial proximity matching method is used to identify residential and commuting activities. An XGBoost model is built with household and personal attributes, activity chain, and spatiotemporal characteristics. The model is employed to classify the types of non-home-non-work activities. Thirdly, we cut trip trajectory into trip slices and extract features referred to motion, trip, and GIS data, and infer travel mode by the XGBoost model. The proposed trip chain inference method is validated with an experimental dataset collected by travel survey system. Results show that the precision and recall of anchors extraction are 96.7% and 96.4%, respectively; the precision of identifying trip ends and public transport transfer nodes are 97.6% and 91.8%, respectively. For trip purpose inference, the model achieves accuracies of 100% for home, 89.8% for work/education, and 87.6% for non-home-non-work activities. For travel mode inference, each mode gets an accuracy of over 90%, and the comprehensive accuracy reaches 95.0%. This paper provides a method of mining trip chains from GPS trajectory to support the application of smartphone-based household travel surveys in the real world.