Generalized-CVO: 10x faster point cloud registration with Riemannian optimization

Researchers from Toyota Research Institute and MIT propose Generalized-CVO, a new point cloud registration method that does not require explicit point correspondences. The key innovation is representing point clouds as continuous functions using reproducing kernel Hilbert space (RKHS) embeddings with anisotropic kernels. These kernels encode local surface geometry, allowing better alignment along surface normals while relaxing constraints along tangential directions—a natural fit for real-world LiDAR and RGB-D data.

To solve the resulting optimization, the team introduces a second-order Riemannian optimization scheme with approximate Hessians, achieving a 10x speedup compared to prior first-order solvers. Experiments on driving datasets show over 55% reduction in both translational and rotational drift in challenging feature-sparse environments. On object registration benchmarks, Generalized-CVO outperforms ICP-based methods and provides notable gains when refining global initializations, especially under moderate misalignment. The method is particularly promising for autonomous vehicles and robotics requiring real-time, reliable 3D tracking.