Implement Point Feature Histogram
Point set registration and alignment of point clouds is used across numerous applications including medical imaging, virtual and augmented reality, autonomous driving, and reconstruction of 3D objects from multiple laser scans. Finding alignments not only accurately but quickly is critical for the merging of multiple data sets into a globally consistent model and mapping new data to old data. 3D data from LiDARs and RGB-D cameras often have various densities, noise levels, and initial poses, all of which present particular challenges with aligning 3D points with speed and accuracy. A simplistic approach to point set registration is the Iterative Closest Point algorithm using Euclidean distance to find the point correspondences between a source and a target point cloud. We present an alternative approach to point set registration that is a significant improvement of this method by instead using pre-computed point feature histograms, which describe more about the curvature of the surface around the given point. As a result, our solution is pose invariant, robust to noise, and not sensitive to different sampling densities.