The two most common display technologies used in augmented reality head-mounted displays are optical see-through and video see-through. In this paper I demonstrate a third alternative: reconstructive see-through. By using a commodity depth camera to construct a dense 3D model of the world and rendering this to the user, distracting latency and position discrepancies between real and virtual objects can be reduced.

Existing head-tracking systems all suffer from various limitations, such as latency, cost, accuracy, or drift. I propose to address these limitations by using depth cameras and existing 3D reconstruction algorithms to simultaneously localize the camera position and build a map of the environment, providing stable and drift-free tracking. This method is enabled by the recent proliferation of light-weight, inexpensive depth cameras. Because these cameras have a relatively slow frame rate, I combine this technique with a low-latency inertial measurement unit to estimate movement between frames. Using the generated environment model, I further propose a collision avoidance system for use with real walking.