We propose a probabilistic approach for compensating motion artifacts in 3D in vivo SD-OCT (spectral-domain optical coherence tomography) tomographs. Subject movement causing axial image shifting is a major problem for in vivo imaging. Our technique is applied to analyze the tissue at percutaneous implants recorded with SD-OCT in 3D. The key challenge is to distinguish between motion and the natural 3D spatial structure of the scanned subject. To achieve this, the motion estimation problem is formulated as a conditional random field (CRF). For efficient inference, the CRF is approximated by a Gaussian Markov random field. The method is verified on synthetic datasets and applied on noisy in vivo recordings showing significant reduction of motion artifacts while preserving the tissue geometry.