In the domain of computer vision, registration is the process of calculating the transformation between a known object, called a marker, and a camera which is viewing it. Registration is the foundation for a number of applications across a range of disciplines such as augmented reality, medical imaging and robotic navigation. In the set of two dimensional planar markers, there are two classes: (1) fiducial, which are designed to be easily recognisable by computers but have little to no semantic meaning to people, and (2) natural features, which have meaning to people, but can still be registered by a computer. As computers become more powerful, natural feature markers are increasingly the more popular choice; however there are still a number of inherent problems with this class of markers. This thesis examines the most common shortcomings of natural feature markers, and proposes and evaluates solutions to these weaknesses. The work starts with a review of the existing planar registration approaches, both fiducial and natural features, with a focus on the strengths and weaknesses of each. From this review, the theory behind planar registration is discussed, from the different coordinate systems and transformations, to the computation of the registration transformation. With a foundation of planar registration, natural feature registration is decomposed into its main stages, and each stage is described in detail. This leads into a discussion of the complete natural feature registration pipeline, highlighting common issues encountered at each step, and discussing the possible solutions for each issue. A new implementation of natural feature registration called the Optical-flow Perspective Invariant Registration Augmentation (OPIRA) is proposed, which provides vast improvements in robustness to perspective, rotation and changes in scale to popular registration algorithms such as SIFT, SURF, and the Ferns classifier. OPIRA is shown to improve perspective invariance on average by 15% for SIFT, 25% for SURF and 20% for the Ferns Classifier, as well as provide complete rotation invariance for the rotation dependent implementations of these algorithms. From the investigation into problems and potential resolutions at each stage during registration, each proposed solution is evaluated empirically against an external ground truth. The results are discussed and a conclusion on the improvements gained by each proposed solution and the feasibility of use in a real natural feature registration application is drawn. Finally, some applications which use the research contained within this thesis are described, as well as some future directions for the research.