Abstract:
Registration of multi-view range data containing low-frequency warps is prone to accumulation errors and misalignments of overlapping regions. To this end, a multi-view non-rigid registration algorithm is proposed based on the multiple thin-plate splines. Firstly, the correspondences between each pair of overlapping views are established via a local iterative closest point(ICP) rigid registration algorithm. Then, using the global energy formula in terms of the multiple thin-plate splines, a thin-plate spline transformation is computed for each view by minimizing the sum of squared distances between the corresponding points. Finally, the optimized thin-plate spline transformations are applied to all of views. By introducing the initial point positions of each view as constraints, the registered range data can keep their original shapes as much as possible. To accelerate the solution of optimization, each thin-plate spline is iteratively solved and the radial basis functions are progressively added in the optimization process. Experimental results showed that the proposed algorithm is accurate and efficient, and can effectively eliminate the accumulation errors and improve the alignments of overlapping regions.