An approximate positions-of-people determination means determines, for each partial area, approximate positions of people on an image on the basis of information about the partial area. A states-of-people estimation means, in addition to predicting the current states of particles from the states of particles indicating past states of people and stored in a storage means, adds new particles and evaluates the likelihood of the states of predicted particles and the added particles on the basis of an observation model generated for the observed people to update the weights of the particles, performs particle re-sampling at a probability proportional to the weights of those particles, and outputs the states of the obtained particles as a flow of people.