An improved "Feature Space Projection Method" for long-term traffic information forecasting is proposed. The traditional method can deal with various temporal factors which traffic condition depends on, such as days, seasons, and vacations. It also achieves accurate forecasting with a low amount of calculation. An internet service using this method provides drivers with the traffic information forecast. However, it is difficult to apply this method to a personal computer for home use and in-car navigation unit because the size of a forecasting database is large. A method is introducedthat decreases the size of the forecasting database greatly by performingforecast process in the feature space shared by several links keeping theaccuracy of forecast data. In case of the Japanese nationwide database, it decreased the size of the database to one-tenth of the usual one derivedby regression analysis. Using this new method, the total size of Japaneseforecasting database is decreased to 820MB. For the covering abstract seeITRD E134653.
Abstract