Augmented Reality (AR) applications rely on efficient and robust methods of tracking. One type of tracking uses dense 3D point data representations of the object to track. As opposed to sparse, dense tracking approaches are highly accurate and precise by considering all of the available data from a camera. A major challenge to dense tracking is that it requires a rough initial matching and mapping to begin. A matching means that from a known object, we can determine the object exists in the scene, and a mapping means that we can identify the position and orientation of an object with respect to the camera. Current methods to provide the initial matching and mapping require the user to manually input parameters, or wait an extended amount of time for a brute force automatic approach.

The research presented in this thesis develops an automatic initial matching and mapping for dense tracking for AR, facilitating natural AR systems that track 3D objects. To do this, an existing offline method for registration of ideal 3D object point sets is proposed as a starting point. The method is improved and optimized in four steps to address the requirements and challenges for dense tracking in AR with a noisy consumer sensor. A series of experiments verifies the suitability of the optimizations, using increasingly large and more complex scene point clouds, and the results are presented.