GVP-Net: An Attention-based Keypoint Generation Network for LiDAR Point Cloud Registration

With the rapid advancement of digital cities and machine vision, 3D reconstruction techniques are playing an increasingly vital role in autonomous driving, robotic path planning, and intelligent navigation. However, the inherent complexity of outdoor environments and the uncertainties introduced during data acquisition present significant challenges. Factors such as environmental noise, sparse and uneven point cloud distributions, and low overlap rates hinder effective feature extraction and compromise the accuracy and efficiency of reconstruction. The network proposed integrates an attention mechanism for dynamically generating keypoints and introduces a feature cosine similarity constraint to enhance point cloud matching. Unlike traditional methods that search among existing points, the dynamically generated keypoints are determined based on the matching probabilities of a set of candidate points, effectively mitigating the influence of outliers. Building on this, multi-scale features of the keypoints are extracted, fused, and further refined using cosine similarity constraints to improve the accuracy of correspondence matching. Through evaluations on outdoor LiDAR point cloud dataset, the proposed network shows enhanced accuracy and efficiency over existing approaches. On the Kitti dataset, the method demonstrates robust registration performance, achieving a relative translation error of 5.6 cm, a relative rotation error of $0.29^\circ$, and a registration recall of 99.7%. These outcomes exceed those of contemporaneous methods by minimizing both translation and rotation errors.