An output feedback adaptive control scheme is proposed for a class of nonlinear systems with unmodeled dynamics and dynamic uncertainties as well as the unmeasured states. Neural networks are used to approximate the unknown continuous functions, and the unknown system states are reconstructed by using K-filters. By the novel description to unmodeled dynamics, the dynamic signal used to dominate the unmodeled dynamics in the existing literature is avoided. Based on the dynamic surface design method, the estimations of the unknown continuous functions produced in the course of theoretical analysis are removed. The complexity of the design is reduced. By using the method of the Lyapunov function, all the signals in the closed-loop control system are proved to be bounded semi-globally, uniformly and ultimately. Simulation results illustrate the effectiveness of the proposed approach.