To solve the problem of low precision characters and non-Gaussian distribution noise of the micro-satellite attitude determination system, UPF(Unscented particle filter) was proposed. Based on the different characters between the UKF(Unscented Kalman filter) and PF(particle filter), UPF uses the UKF to generate sophisticated proposal distributions. UPF can not only avoid the limitation of the EKF(extended Kalman filter) and the UKF which only apply to Gaussian distributions but also avoid the limitation of the standard PF which can not include the new measurements. The MEMS(micro electronic mechanical system) gyros and CMOS APS(complementary metal oxide semiconductor active pixel sensors) star sensor were applied as attitude sensors, the method of attitude determination used the error quaternion as the attitude parameter, presented a UPF attitude estimator. The attitude determination filter was constructed and the stochastic noise was adopted from the MEMS gyro. Then the UPF characters were analyzed. The simulation results showed that the suggested method can improve the system performance. The determination accuracy was higher than the UKF while the number of the particle was lesser than standard PF.