Current work

Intelligent selection of points refers to the method by which mesh points are selected for comparison in the correspondence matching phase of the registration algorithm. Currently, points are selected at random with the goal of evenly distributing the samples over the surface. Although this naive approach works well for object modeling and recognition, where the surfaces abound with interesting features. Unstructured terrain typically contains large, featureless regions, which do not contribute to registration. Thus, another reasonable criterion for point selection would be an estimate of the degree a point will contribute to the registration. By incorporating this additional criterion into the point selection process, we intend to improve the robustness of the current algorithm.

We are also studying the estimation of required overlap between views. We would like to determine, based on a given data set, how much overlap is necessary before registration will fail. By looking at the way the surface points constrain registration for a particular area of overlap, we can estimate in advance the probability of successful registration from that view. This capability is a first step to the long-term goal of an integrated planning and terrain modeling system.

Example of intelligent selection of points: Some sample points (top) are underconstrained. They are similar to other points spread over a wide area of the surface. Some points (middle) are well constrained. No other points are similar. And some points (bottom) are constrained only in some directions.