We report a new method for fixed-node diffusion quantum Monte Carlo calculation in order to improve the time step-size convergence and the fluctuation problems associated with nodes of wave functions. It adopts floating spherical Gaussian orbitals instead of Slater-type orbitals to guide random walkers (psips). They establish harmonic velocity fields under which psips undergo random walks which are simulated exactly according to the Ornstein-Uhlenbeck process. The nodes of this guiding function are used to guide psips in a novel way; as psips approach to a node, they are swept away therefrom. Explicit enforcement of a sink boundary condition is avoided to some extent, and the behavior of the random walks near nodes is efficiently simulated. The method is illustrated by applying to the ground-state energies of H-2, LiH, and HE The results are excellent compared to previous calculations.