In this article a robust thermal face recognition methodology based on the use of local interest points and descriptors, is proposed. The methodology consists of the following stages: face segmentation, vascular network detection, wide baseline matching using local interest points and descriptors, and classification. The main contribution of this work is the use of a standard wide baseline matching methodology for the comparison of vascular networks from thermal face images. The proposed methodology is validated using a database of thermal images. This work could be of high interest for HRI applications related with the visual recognition of humans, as the ones included in the RoboCup @Home league, because the use of thermal images may overcome limitations such as dependency on illumination conditions and facial expressions.