The stellar/inertial integrated navigation system, which combines the inertial navigation system (INS) and the star tracker, can restrain the accumulated INS errors. In the traditional loosely coupled stellar/inertial integration method, the star tracker needs to observe more than two navigation stars on an image for attitude determination and to use the attitude information as the observation to estimate the systematic errors of the INS. However, under strong background radiation conditions, the star number in the field of view (FOV) usually drops below 3; thus, the loosely coupled method fails to work. To overcome this difficulty, an improved tightly coupled stellar/inertial integration method based on the observation of the star centroid prediction error (SCPE) is proposed in this paper. It calculates the difference between the extracted star centroid and the predicted star centroid, namely, the SCPE, as the observation and then estimates the INS errors with a Kalman filter. Numerical simulations and ground experiments are conducted to validate the feasibility of the tightly coupled method. It is proved that the proposed method, which makes full use of all star observation information, can improve the navigation accuracy compared with the loosely coupled method and is more robust when there are not enough stars in the FOV.