One of the most challenging issues in task-related fMRI data analysis consists of deriving a meaningful functional brain parcellation. The joint parcellation detection estimation (JPDE) model addresses this issue through an automatic inference of the parcels directly from fMRI data. However, for doing so, the number of parcels needs to be fixed a priori and an appropriate initialization for the mask parcellation must be provided too. Hence, this difficult task generally depends on the subject. In this paper, an automatic model selection approach is proposed to overcome this limitation at the subject-level. Our approach relies on a non-parametric Bayesian approach that estimates the number of parcels online using a Dirichlet process mixture model combined with a hidden Markov random field. The inference is carried out using a variational expectation maximization strategy. As compared to a standard model selection approach in the original JPDE framework, our non-parametric extension appears more efficient in terms of computational time and does not require finely tuned initialization. Our method is first validated on synthetic data to demonstrate its robustness in selecting the right model order and providing accurate estimates for the parcellation, the hemodynamic response function (HRF) shapes and the activation maps. The method is then validated on real fMRI data in two regions of interest (ROIs): right motor and bilateral occipital ROIs. The results show the ability of the proposed method to aggregate parcels with similar behaviour from a hemodynamic point of view, while discriminating them from other parcels having different hemodynamic properties. The HRF estimates of the dfferent hemodynamic territories obtained with our approach are close the the canonical HRF shape in both the right motor and the bilateral occipital cortices. The discrimination power of the proposed approach is increased compared to its ancestors where the results on real data show its ability to discriminate HRF profiles with different Full Width at Half Maximum (FWHM). The robust performance of detecting the elicited task-related activity is confirmed by comparing the neural response level estimates obtained using our approach with those obtained using the joint detection estimation (JDE) model.