A storage unit stores a model defining a position or a locus of a feature point of an occupant in each specific action. An action estimation unit compares the feature point with each of the models to detect an estimated action. A detecting unit detects that a specific action is being performed as a definite action. A first generating unit generates a new definite model corresponding to the definite action by modifying a position or a locus of the feature point according to an in-action feature point when the definite action is being performed. A second generating unit generates a new non-definite model using the in-action feature point according to a correspondence between the feature point in the definite action and the feature point of a non-definite model other than the definite model. An update unit updates the definite action model and the non-definite action model.