Some modern spacecraft missions require precise knowledge of the attitude, obtained from the ground processing of on-board attitude sensors. A traditional 6-state attitude determination filter, containing three attitude errors and three gyro bias errors, has been recognized for its robust performance when it is used with high quality measurement data from a star tracker for many past and present missions. However, as higher accuracies are required for attitude knowledge in the missions, systematic errors such as sensor misalignment and scale factor errors, which could often be neglected in previous missions, have become serious, and sometimes, the dominant error sources. The star tracker data have gaps and degradation caused by, for example, the Sun and Moon blocking in the filed of view and data time tag errors. Thus, attitude determination based on the gyro data without using the star tracker data is inevitably required for most missions for the period when the star tracker is unable to provide accurate data. However, any gyro-based attitude errors would eventually grow exponentially because of the uncorrected systematic errors of gyros and the uncorrected gyro random noises. An improved understanding of the gyro random noise characteristics and the estimation of the gyro scale factor errors and gyro misalignments are necessary for precise attitude determination for some present and future missions. The 6-state filters have been extended to 15-state filters to estimate the scale factor and misalignment errors of gyros especially during a high-slew maneuver and the performance of theses filters has been investigated. During a starless period, the inevitable drift of the EKF solutions, which are caused by the uncorrected gyro’s systematic errors and the gyro random noises, can be replaced with the batch solutions, which are less affected by the data gap in the star tracker. Power Spectral Density and the Allan Variance Method are used for analyzing the gyro random noises in both ICESat and simulated gyro data, which provide better information about the process noise covariance in the attitude filter. Both simulated and real data are used for analyzing and evaluating the performances of EKF and batch algorithms. / text