Head motion is one of the major issues in neuroimaging. With the introduction of MR-PET scanners, motion parameters can now be estimated from two independent modalities acquired simultaneously. In this work, we propose a new data-driven method that combines MR image registration and PET data driven approach to model head motion during the complete course of MR-PET examination. Without changing the MR-PET acquisition protocol, the proposed method provides motion estimates with a temporal resolution of ~2 secs. Results on a phantom dataset show that the proposed method can significantly reduce motion artefact in brain PET images and improve image sharpness compared with the MR based methods.