In this paper, we explore methods of robustly extracting fiducial facial feature points - an important process for numerous facial image processing tasks. We consider various methods to first detect face, then facial features and finally salient facial feature points. Colour-based models are analysed and their overall unsuitability for this task is summarised. The bulk of the report is then dedicated to proposing a learning-based method centred on the Viola-Jones algorithm. The specific difficulties and considerations relating to feature point detection are laid out in this context and a novel approach is established to address these issues. On a sequence of clear and unobstructed face images, our proposed system achieves average detection rates of over 90%. Then, using a more varied sample dataset, we identify some possible areas for future development of our system.