Signed Distance Registration for Depth Image Sequence

New depth camera technology has potential to make a significant impact on computer systems interaction with 3D objects; yet, it is currently limited due to its poor noise and resolution characteristics. In this thesis we propose to use depth camera’s strongest characteristic, its video rate capture speeds, to overcome these limitations. Previous work to utilize sequences of depth images used 2D super-resolution techniques to combine chunks of depth images that are close in time in order to increase the resolution and noise characteristic. This technique took advantage of the consistency of the scene with respect to small changes in viewpoint. But, while increasing the resolution and decreasing unbiased noise, this algorithm increased biased noise. Thus, we are motivated to consider an algorithm that can first register all depth images to a common 3D coordinate system and then utilize a 3D superresolution technique, known as surface reconstruction, to increase the resolution and decrease both biased and unbiased noise. This thesis considers the first part of this problem, which is the registration of a sequence of depth images to a common coordinate frame. Previous registration methods were developed for high resolution, low noise point clouds and perform poorly for noisy sequences of depth data. Thus, from our analysis of ideal signed distance functions, we propose a new method for finding the closest point to the surface given a signed distance function and its gradient. Utilizing Implicit Moving Least Squares (IMLS) and our analysis, we propose a new algorithm that computes the registration of a set of points to the surface defined by the IMLS function of another set of points. We also propose a grid based implementation that allows for bounded computations per time step. Our results demonstrate that the proposed algorithm is more robust in the presence of realistic depth noise.