Toyota Research Institute published Generalized-CVO, a correspondence-free local point cloud registration method using second-order Riemannian optimization with approximate Hessians and reproducing kernel Hilbert space embeddings1. The method encodes local geometry through point-wise anisotropic kernels and achieves up to 10x speedup over first-order correspondence-free RKHS-based solvers. On a LiDAR tracking task in driving environments, Generalized-CVO reduced translational and rotational drift by 55% in feature-sparse conditions and demonstrated improved robustness over ICP-based methods on object registration benchmarks.
Toyota Research Publishes Generalized-CVO for Point Cloud Registration
Toyota Research Institute's Generalized-CVO achieves 10x speedup and 55% drift reduction in correspondence-free point cloud registration using…