ECG imaging is a non-invasive technique of characterizing the electrical activity and the corresponding excitation conduction of the heart using body surface ECG. The method may provide great opportunities in the planning of cardiac interventions and in the diagnosis of cardiac diseases. This work introduces an algorithm for the imaging of transmembrane voltages that is based on a Kalman filter with an augmented measurement model. In the latter, a regularization term is integrated as additional measurement. The filter is trained using a-priori-knowledge from a simulation model. Two effects are investigated: the influence of the training data on the reconstruction quality and the representation of a-priori knowledge in the trained covariance matrices. The proposed algorithm shows a promising quality of reconstruction and may be used in the future to introduce generic physiological knowledge in solutions of cardiac source imaging.