A bi-level programming model to reconstruct origin-destination (OD) demand by using density as the observed variable under congested network is proposed. The upper-levels minimize the errors on the estimated values and observed values, and the lower-levels are user equilibrium model. For a bi-level programming model, KKT condition method is adopted, it is transformed into a mathematical program with equilibrium constraints which is easier to solve, and then Scholtes relaxation method is used to solve the transformed model. The numerical results show that, using density as the observed variable is better than using flow in the OD reconstruction problem under congested network. Meanwhile, for solving method of bi-level programming model, the method of transforming KKT condition into single-level is superior to the upper-lower alternate algorithm.