In this paper, unscented Kalman filter (UKF) is studied to estimate the state of Dead reckoning (DR) and GPS integrated navigation system. The positioning error of DR system mainly comes from two factors, the azimuth error and odometer scale factor error. Conventional model of DR/GPS integrated navigation chooses acceleration, position and velocity as observation states, and azimuth error is not estimated, which is one of the key error sources. A new error model of DR/GPS system is adopted that includes azimuth error as a state, which makes it possible to estimate both of the two major error sources. UKF directly approximates the probability density distribution of random variable and avoids the linearization of nonlinear function, which improves the filtering precision. UKF is used to implement an improved DR/GPS system. Road test has been conducted to prove the effectiveness of the scheme.