Using k-d trees for robust 3D point pattern matching

We propose a new method for matching two 3D point sets of identical cardinality with global similarity but locally nonrigid distribution. This problem arises from marker-based optical motion capture systems. The point-sets are extracted from similar design poses of two subjects with underlying nonrigidity and possible distribution discrepancies, one being a model set (manually identified) and the other representing observation of another subject, to be matched to the model set. There exists neither a single global scale, nor an affine transformation between the point-sets. To establish the goal of a one-to-one for identification, we introduce a k-dimensional tree based method, which is well adapted and robust to such data, typically with distribution errors due to underlying subject nonrigidity. First, we construct a k-d tree for the model set. Then a similarity k-d tree of the data set is constructed following the structure information embedded in the model tree. Matching sequences of the two point sets are generated by traversing the identically structured trees. Experimental results confirm that this method is applicable for robust spatial matching of sparse point sets under nonrigid distortion.