Point pattern matching for articulated or multiple objects

An efficient point pattern matching algorithm for articulated and multiple objects is presented in this paper. A local to global strategy is adopted to get three layered matches starting from an initial partition of the point set using clustering method. Firstly, initial match with three point correspondences is obtained through local neighborhood and exhaustive search. Then, central match is expanded from initial match by alternating between next point matching and transformation update. Finally, ambiguous boundary points are matched and classified into their correspondent parts using the estimated alignment transformations, and missing or extra points are detected and rejected as outliers. Experiments on real images present satisfying results.