Title: Destination prediction by trajectory distribution-based models
Authors: Brendan Guillouet - Institut de Mathematiques de Toulouse (France) [presenting]
Jean-Michel Loubes - University of Toulouse (France)
Philippe Besse - Institut de Mathematiques de Toulouse (France)
Francois Royer - Datasio (France)
Abstract: A data-driven methodology to predict the final destination of vehicle trips based on their initial partial trajectories is introduced. First, a clustering of trajectories is produced using hierarchical clustering and based on a new distance between trajectories, the Symmetrized Segment-Path Distance, (SSPD). The clusters obtained describe the main patterns of the traffic flow based on the drivers' usage. Locations within each of these pattern are then modeled by a mixture of $2d$ Gaussian distributions. Hence, a data driven grid is learned based on the behaviors of the drivers. Finally, this model is used to predict the final destination of a new trajectory based on their first locations with a two step procedures: the new trajectory are first assigned to the clusters it belongs the most likely. Secondly, the final destination is predicted using characteristics from trajectories inside these clusters. This methodology has been successfully tested on two different datasets, assessing its capacity to be adapted to different networks. One of these datasets comes from the ECML/PKDD 15: Taxi Trajectory Prediction Kaggle challenge. Our final results produce promising result taking into account that our model can be re-used directly for a different test dataset, and can also be used to predict the destination during trajectory completion, without requiring a new training.