Adaptive Downsampling and Refinement for Fast Iterative Closest Point

Iterative closest point (ICP) and its variants rep resent a fundamental technique for rigid registration between two point sets, with wide applications ranging from robotics to 3D reconstruction. The primary limitations of ICP include slow convergence due to its linear convergence rate, along with the per-iteration nearest-neighbor search, which scales linearly with source cloud size and dominates runtime on dense datasets. To address these issues, this paper proposes a modified ICP method that achieves fast convergence while maintaining robust alignment. Specifically, we introduce an adaptive threshold-based downsampling technique that selectively reduces the number of source points processed during early iterations, decreasing the computational cost for each iteration. This is followed by a full cloud refinement stage designed to recover alignment accuracy and mitigate any degradation introduced by the downsampling phase. Simulation results demonstrate that the proposed method achieves notable reductions in computational time across the majority of benchmark datasets, including Stanford mesh models and LiDAR point clouds, while maintaining comparable registration accuracy relative to the baseline method.