Vehicle velocity estimation is essential for the enhancement of safety, especially for yaw stabilization systems. This paper proposes a sliding mode observer (SMO) for longitudinal and lateral velocities estimation on non-flat road based on Dugoff's tire model and nonlinear vehicle dynamics. The road grade and bank angles are considered as unknown inputs, which are reconstructed from their corresponding sliding modes by using the SMO. The stability of the SMO and robustness with respect to additive disturbances are analyzed with the help of input-to-state stability (ISS) theory. The SMO is evaluated in comparison with a nonlinear observer without road angle compensation, and its performance and robustness is validated using vehicle simulator software veDYNA under a variety of maneuvers and road conditions.