<p>Point cloud registration is a fundamental technology in 3D data processing, with critical applications in autonomous systems, robotics, and large-scale environmental mapping. Significant progress has been made in correspondence-based point cloud registration techniques. However, the accuracy of registration models is often significantly compromised by environmental occlusions, sensor limitations, and inherent algorithmic constraints. This paper presents a robust estimator for point cloud registration, Progressive Local Optimization–Improved Progressive Sample Consensus (PLO-IPROSAC). PLO-IPROSAC introduces a quality function based on keypoint geometric consistency, which uniformly guides both global sampling and local optimization processes, thereby significantly enhancing efficiency. Building on this, we first design a Progressive Local Optimization (PLO) method, which prioritizes high-quality hypotheses and employs a more refined Gaussian loss function for model evaluation. Under noisy and occluded conditions, PLO achieves lower errors and higher efficiency compared to mainstream local optimization methods. Experiments demonstrate that, compared to RANSAC and its variants, PLO-IPROSAC not only achieves state-of-the-art accuracy but also reduces the minimum number of iterations by 90% relative to the second-best performing algorithm. On standard datasets including Bologna, Kinect, and UWA, the proposed method exhibits superior accuracy and robustness.</p>