Registration of 3D point clouds using a local descriptor based on grid point normal

The coarse-to-fine method is the prime technology for point cloud registration in 3D reconstruction. Aimed at the problem of low accuracy of coarse registration for the partially overlapping point clouds, a novel, to the best of our knowledge, 3D local feature descriptor named grid normals deviation angles statistics (GNDAS) for aligning roughly pairwise point clouds is proposed in this paper. The descriptor is designed by first dividing evenly the local surface into some grids along the x axis and y axis of the local reference frame, then making the statistics of the deviation angles of normals at grid points. Based on the correspondences generated by matching descriptors and a transformation estimation method, the initial registration result is obtained. The coarse registration result is used as the initial value of the iterative closest point algorithm to achieve the refinement of the registration result. Experimental comparisons on two public datasets demonstrate that our GNDAS descriptor has high descriptiveness and strong robustness to noise at low level and varying mesh resolution. The registration results also confirm the superiority of our registration approach over previous versions in accuracy and efficiency.