一种用于大旋转角和严重非刚性形变点集配准的新型局部特征描述符

A Novel Local Feature Descriptor for Point Set Registration with Large-Angle Rotations and Severe Non-Rigid Deformations

摘要: 点集配准是计算机视觉中常见的研究方向，其目的是对两个点集执行空间对齐，目前已被广泛应用于点云和图像配准中。常规刚性点集配准问题的主要挑战在于求解出两个点集之间的对应关系，并以此估算出两者之间的旋转平移矩阵。而在实际应用场景中，提取出来的两个点集之间的刚性变换偶尔会出现大角度旋转，而且两点集间还会出现非刚性形变，从而给配准问题带来更大的挑战。点集局部特征描述符可以清晰刻画出点集的局部几何结构并提取出点集之间的结构差异，更有利于后续的配准过程，因此被广泛应用于现有的各类型点集配准算法中。然而现有的局部特征描述符主要集中于解决常规旋转平移问题和非刚性形变问题，极少考虑解决大角度旋转和非刚性形变同时作用下的耦合效应。为此本文提出一种新颖的点集局部特征描述符，该描述符基于局部点集间的欧氏距离、三角形角度特征和伪对应关系而构建，随后本文将所提特征描述符嵌入至经典的非刚性点集配准算法中，并开展实验验证。对比实验结果表明，嵌入所提特征描述符之后的配准算法在处理非刚性形变、大角度旋转以及两者共同作用下的配准问题中，取得了最优的配准效果。

Abstract: Point set registration is a prevalent research direction in computer vision, which aims to achieve spatial alignment between two point sets and has been extensively applied in point cloud and image registration tasks. The core challenge of conventional rigid point set registration lies in establishing correspondences between the two point sets and estimating the rotation-translation matrix based on these correspondences. However, in practical scenarios, rigid transformations between the extracted point sets may occasionally involve large-angle rotations, and non-rigid deformations often exist between the two sets, thereby introducing greater challenges to the registration problem. Local feature descriptors for point sets can explicitly characterize the local geometric structures of point sets and capture structural discrepancies between them, which is conducive to subsequent registration processes and thus widely adopted in various existing point set registration algorithms. Nevertheless, existing local feature descriptors primarily focus on addressing standard rotation-translation and non-rigid deformation issues, with minimal consideration given to the coupled effects induced by the simultaneous presence of large-angle rotations and non-rigid deformations. To address this gap, this paper proposes a novel local feature descriptor for point sets, which is constructed based on Euclidean distances between local points, triangular angle features, and pseudo-correspondences. Subsequently, the proposed feature descriptor is embedded into a classic non-rigid point set registration algorithm, and experimental validations are conducted. Comparative experimental results demonstrate that the registration algorithm integrated with the proposed descriptor achieves optimal performance in handling registration problems involving non-rigid deformations, large-angle rotations, and their combined effects.