Preliminary results of a new method for range view registration are presented. The method incorporates the LeRP Algorithm, which is a deterministic means to approximate subgraph isomorphisms. Graphs are formed that describe prominent, isolated, scene features. Graph matching then provides the scene-to-scene correspondence necessary to compute a transformation between sets of range data. A graphical representation is used that is invariant with respect to sensor standoff. This permits not only the translation and rotation to be determined, but also scale differences. Results from 10 range images of real sensor data indicate that a reasonable tradeoff between speed and accuracy is achievable. A mean rotational error of 0.53 degrees was found for the 320x240 range images. In the test trials, the mean and maximum compute time was found to be 1.4 and 1.5 seconds, respectively, for the registration calculations. These tests were run on a 366 MHz PC. The greatest challenge to this approach is the stable localization and characterization of features in the range images - when computed via simple, rapid, techniques. This type of approach may have best application in situations without a high degree of clutter and occlusion. Tests presented herein focus on the registration of aerial range data, which appears to be a good candidate application.
