Objects. In functional stereotactic neurosurgery, one of the cornerstones upon which the success and the operating time depends is an accurate targeting. The subthalamic nucleus (STN) is the usual target involved when applying Deep Brain Stimulation (DBS) for Parkinson's disease (PD). Unfortunately, STN is usually not clearly visible in common medical imaging modalities, which justifies the use of atlas-based segmentation techniques to infer the STN location. Materials and Methods. 8 bilaterally implanted PD patients were included in this study. A three-dimensional T1-weighted sequence and inversion recovery T2-weighted coronal slices were acquired pre-operatively. We propose a methodology for the construction of a ground truth of the STN location and a scheme that allows both, to perform a comparison between different non-rigid registration algorithms and to evaluate their usability to locate the STN automatically. Results. The intra-expert variability in identifying the STN location is 1.06 ± 0.61 mm while the best non-rigid registration method gives an error of 1.80 ± 0.62 mm. On the other hand, statistical tests show that an affine registration with only 12 degrees-of-freedom is not enough for this application. Conclusions. Using our validation-evaluation scheme we demonstrate that automatic STN localization is possible and accurate with non-rigid registration algorithms.