In this paper, the well-known multiplicative extended Kalman filter (MEKF) is re-investigated for attitude estimation using vector observations. From the Lie group theory, it is shown that the attitude estimation model is group affine and its error state model should be trajectory-independent. Moreover, with such trajectory-independent error state model, the linear Kalman filter is still effective for large initialization errors. However, the measurement model of the traditional MEKF is dependent on the attitude prediction, which is therefore trajectory-dependent. This is also the main reason why the performance of traditional MEKF is degraded for large initialization errors. Through substitution of the attitude prediction related term with the vector observation in body frame, a trajectory-independent measurement model is derived for MEKF. Meanwhile, the MEKFs with reference attitude error definition and with global state formulating on special Euclidean group have also been studied, with main focus on derivation of the trajectory-independent measurement models. Extensive Monte Carlo simulations and field test of attitude estimation implementations demonstrate that the performance of MEKFs can be much improved with trajectory-independent measurement models.