A new data fusion algorithm for point cloud registration

In recent years, numerous methods based on Gaussian process algorithms, weighted-least-square algorithms and machine learning algorithms have been proposed for three-dimensional point cloud registration. However, these algorithms have often been tested on point clouds of similar sizes and point densities, frequently sharing similar initial alignments and orientations. In this paper, we propose a new algorithmic pipeline for registering two point clouds of different sizes and point densities, that do not share initial alignment. This algorithm is used to register one dataset that is dense and small with one that is sparse and large; the former representing a region in the latter dataset. Our algorithm, firstly, sets the large point cloud as a reference and segments it into subsections ("sub-clouds") of the exact size of the small point cloud. Then, the algorithm compares the geometrical similarities between each sub-cloud and the small point cloud: both are further partitioned into layers along an arbitrary axis, with each layer again being partitioned into identical voxels. The number of points contained in each voxel is divided by the total number of points in each point cloud (i.e. converted to a percentage of points). If the percentages of points in a pair of corresponding voxels in both point clouds are similar, the pair is considered to be matched. Then, if > 90 % of the total number of voxels in this layer are matched, this pair of layers are labelled as matched. If > 90 % of the total number of layers are matched, this sub-cloud is regarded as successfully matched to the small point cloud. Finally, the small point cloud is registered to the location of the large point cloud in the reference coordinate system (the coordinate system of the big point cloud). Our algorithm has been tested with synthetic datasets, showing initial success. In future work, the pipeline will be tested on real measurement data.