In various robotics applications, matching several sets of points (aka point clouds) is an essential component. Typical applications are found in mobile robotics(e.g. localization), industrial robotics (e.g. bin picking), medical robotics(e.g. fracture reduction), and many other areas.
The matching problem is often defined as: Given two point clouds A and B, find a translation vector t and a rotation matrix R resulting in a maximum overlap of the two point clouds.
A well-established approach to matching problems is the Random Sample Consensus(RANSAC) method[1] which randomly generates hypotheses for R and t based on a minimal subset of points from A and B and then validates them. Winkelbach[2] proposed a technique, called RANSAM, which drastically accelerates the search for hypotheses. Additional speed-ups may be gained by parallelization[3].
The talk shall first give an introduction to the matching problem and then focus on the core ideas behind RANSAM and its extensions.
Sources:
