Hemodynamic Response Function (HRF) estimation in noisy functional Magnetic Resonance Imaging (fMRI) is essential for a better understanding of cerebral activations. Previous works have proposed robust non-parametric estimates of the HRF within a regularized framework [C. Goutte et al., 2000, G. Marrelec et al. 2001]. They are not adapted for event-related paradigms that are either asynchronous (in which the onsets of the conditions are not synchronized with the data), or designed with several kinds of trials, or with several sessions. In this paper, we extend the previous works to these three situations. We introduce temporal prior information on the underlying physiological process of the brain hemodynamic response to accurately estimate the HRF in a Bayesian framework. The proposed unsupervised approach is validated on both synthetic and real fMRI data.