For MEMS-based inertial units, a scheme for horizontal attitude estimation based on nonlinear state observer is proposed by analysing the direction cosine matrix, then the stability of the algorithm is proved. On this basis, a practical scheme based on Kalman filter is designed whose system equation derived from MEMS gyros outputs and measurement constructed by accelerometers outputs in uniform motion. A so-called double model value mechanism is designed to automatically identify whether the measurement information is effective. The algorithm has been evaluated by static dynamic and interference experiments, and the results show that it can successfully produce stable and effective horizontal attitudes and estimation error is within 1°.