Improving short-term travel time prediction for on-line car navigation by linearly transforming historical traffic patterns to fit the current traffic conditions

This research is focused on the problem of travel time prediction for a personal on-line car navigation system. The aim of this study is to improve the short-term travel time prediction quality by creating a dynamic model that utilises the real-time GPS floating car data (from the users of a car navigation system), assuming a static black box model of historical traffic patterns is given as a base. A novel model is introduced for this task. It combines two methods: the first one (applied on the city level) is based on linearly transforming traffic patterns to fit the current traffic conditions by solving a weighted and regularised regression problem. The second method (applied on short road segments) is based on exponential smoothing. The models quality is evaluated through extensive experiments on real data, by measuring the squared prediction error on the chosen observations that has not previously influenced the examined model. The results show significant improvement over the static historical traffic patterns model.