Non-rigid motion estimation is an important task in correction of respiratory and cardiac motion. Usually this problem is formulated in image space via diffusion, parametric-spline or optical flow methods. For these applications non-rigid motion commonly needs to be estimated from cardiac or respiratory states which are highly subsampled in k-space. Image-based registration can be impaired by aliasing artefacts or by estimation from low-resolution images. In this work, we propose a novel non-rigid registration technique directly in k-space based on optical flow. The proposed method is compared against image-based registrations and tested on fully-sampled and highly-accelerated 3D motion-resolved MR imaging.