Range-velocity (RV) map (also called the range-Doppler map) is used in radar to detect targets and estimate the range and velocity of the targets. The conventional range-Doppler processing has its limits in attaining a high-resolution RV map with few samples. Moreover, the conventional range-Doppler processing may cause severe range migration problems and bias in the range estimation. In this paper, we propose to attain a high-resolution and accurate RV map with the sparse Bayesian learning (SBL) framework. Unlike the existing works that use the discrete Fourier transform to build the dictionary for sparse representation, we build a dictionary from a model of the wide-band frequency-modulated continuous-wave (FMCW) radar to avoid the model mismatch problem. This dictionary allows for the effective decomposition of the signal and the full utilization of the structure of the signal. In addition, based on the fast marginal likelihood maximization, we also propose a fast method to reduce the computational burden. Simulations show that the proposed method and the proposed fast method are superior in terms of resolution, accuracy, and maximum unambiguous velocity compared with the existing methods.