Go-ICP: A Globally Optimal Solution to 3D ICP Point-Set Registration

Area of Science:

• Computer Vision
• Robotics
• Computational Geometry

Background:

• Iterative Closest Point (ICP) is a standard algorithm for 3D point-set registration.
• ICP's local optimization nature makes it prone to suboptimal solutions (local minima).
• ICP's performance is highly dependent on initial alignment quality.

Purpose of the Study:

• To introduce the first globally optimal algorithm for Euclidean registration of 3D point-sets.
• To address the limitations of traditional ICP, specifically its susceptibility to local minima.
• To provide a reliable registration method that does not require a good initial guess.

Main Methods:

• Developed Go-ICP, a novel algorithm based on a branch-and-bound (BnB) search scheme.
• Exploited SE(3) geometry to derive new upper and lower bounds for registration error.
• Integrated local ICP within the BnB framework to enhance speed and ensure global optimality.

Main Results:

• Go-ICP guarantees finding the globally optimal solution for rigid 3D point-set registration under the L2 error metric.
• The algorithm demonstrates reliable registration results independent of initialization quality.
• Evaluations confirm the effectiveness of Go-ICP in scenarios demanding optimal solutions.

Conclusions:

• Go-ICP overcomes the local minima problem inherent in traditional ICP.
• The method offers a robust and reliable solution for 3D point-set registration, even with poor initializations.
• Go-ICP is suitable for applications requiring guaranteed optimal registration outcomes.