We present an efficient algorithm for non-rigid registration of two partially overlapping 3D surfaces in which a target surface is a deformed instance of a source surface. The algorithm is implemented in two main phases. In the first phase, the robust algorithm that is used is based on a probability density estimation to find reliable correspondences between the two surfaces. Then, in the second phase, a deformation algorithm is applied for non-rigid registration where the displacement of each point is described by an affine transformation in relation with other points of the same surface and its corresponding point on the other surface. Combined with initial correspondences in the first phase, an effective strategy for optimization of a cost function is carried out to align the two surfaces without using any assumption and user-intervention on the algorithm. We test the robustness of our method by efficiently aligning pairs of surfaces of realistic scan data of human body models.