In dynamic planar imaging, extraction of signals specific to structures is complicated by structures superposition. Due to overlapping, signals extraction with classic regions of interest (ROIs) methods suffers from inaccuracy, as extracted signals are a mixture of targeted signals. Partial volume effect raises the same issue in dynamic tomography. Source separation methods, such as factor analysis of dynamic sequences, have been developed to unmix such data. However, the underlying problem is underdetermined and the model used is not relevant in the whole image. This non-uniqueness issue was overcome by introducing prior knowledge, such as sparsity or smoothness, in the separation model. In practice, these methods are barely used because of the lack of reliability of their results. Previously developed methods aimed to be fully automatic, but efficiency can be improved with additional prior knowledge. Some methods using ROIs knowledge in a straightforward way have been proposed. In this paper, we propose an unmixing method, based on an objective function minimization and integrating these ROIs in a different and robust manner. The objective function promotes consistent solutions regarding ROIs while relaxing the model outside ROIs. In order to reduce user-dependent effects, ROIs are used as soft constraints in a robust way through the use of a distance matrix. Consistency, effectiveness, and robustness to the ROIs selection are demonstrated on a toy example, a highly realistic simulated renography data set and a clinical data set. Performance is compared with the competitive methods.