Parameter-free global hybrid point based range image registration

In order to tackle one of the most difficult tasks in 3D computer vision, range image registration (RIR), this paper proposes a new approach to register two range images of the same object or scenario from different scanning angles. The new method of two dimensional point based boundary searching replaces conventional six dimension methods. This point based approach is applied into hybrid registration paradigm which integrates global searching algorithms with a local alignment tool, Iterative Closest Point (ICP). With this approach, searching algorithms are able to find global optima more efficiently with significantly fewer searching dimensions. Because of using data points as searching variables and taking all of them into consideration, registration algorithms are able to explore the whole space to find the global solution without any parameter constraint. Point based searching approach is successfully implemented on hybrid registration algorithms which use state-of-the-art global searching tools including Simulated Annealing (SA), Differential Evolution (DE) and Particle Swarm Optimization (PSO). The new approach is evaluated in terms of both accuracy and robustness in various experiments on different datasets to prove its superior over the conventional approach.