For pose estimation in visual servoing, by assuming the relative motion over one sample period to be constant, many existing works adopt a linear time invariant (LTI) dynamic model. Since the standard feature point transformation is nonlinear, extended Kalman filtering (EKF) has become popular due to its simplicity. Thus, the problem at hand becomes filtering of an LTI system with a time-varying output matrix. To obtain satisfactory performance, accurate knowledge of the noise covariances is essential. Various methods have been proposed on how to adaptively update their values to improve performance. However, these techniques cannot guarantee the positive semidefiniteness (PSD) of the covariance estimates. In this paper, we propose to apply the autocovariance least-squares (ALS) approach to covariance identification in pose estimation. The ALS approach can provide reliable estimates of the covariance matrices while maintaining their PSD and imposing desired structural constraints. Our tests show that using the covariance estimates from the ALS method in EKF can reduce the average pose estimation error by more than 30% in simulation, and the average position estimation error by about 30% using experimental data, respectively, compared to a hand-tuned EKF.