Effective nearest neighbor search for aligning and merging range images

We describe a novel method, which extends the search algorithm of a k-d tree for aligning and merging range images. If the nearest neighbor point is far from a query, many of the leaf nodes must be examined during the search, which actually will not finish in logarithmic time. However, such a distant point is not as important as the nearest neighbor in many applications, such as aligning and merging range images; the reason for this is either because it is not consequently used or because its weight becomes very small. Thus, we propose a new algorithm that does not search strictly by pruning branches if the nearest neighbor point lies beyond a certain threshold. We call the technique the bounds-overlap-threshold (BOT) test. The BOT test can be applied without recreating the k-d tree if the threshold value changes. Then, we describe how we applied our new method to three applications in order to analyze its performance. Finally, we discuss the method's effectiveness.