Matching Algorithm for 3D Point Cloud Recognition and Registration Based on Multi-Statistics Histogram Descriptors

Abstract

Establishing an effective local feature descriptor and using an accurate key point matching algorithm are two crucial tasks in recognizing and registering on the 3D point cloud. Because the descriptors need to keep enough descriptive ability against the effect of noise, occlusion, and incomplete regions in the point cloud, a suitable key point matching algorithm can get more precise matched pairs. To obtain an effective descriptor, this paper proposes a Multi-Statistics Histogram Descriptor (MSHD) that combines spatial distribution and geometric attributes features. Furthermore, based on deep learning, we developed a new key point matching algorithm that could identify more corresponding point pairs than the existing methods. Our method is evaluated based on Stanford 3D dataset and four real component point cloud dataset from the train bottom. The experimental results demonstrate the superiority of MSHD because its descriptive ability and robustness to noise and mesh resolution are greater than those of carefully selected baselines (e.g., FPFH, SHOT, RoPS, and SpinImage descriptors). Importantly, it has been confirmed that the error of rotation and translation matrix is much smaller based on our key point matching algorithm, and the precise corresponding point pairs can be captured, resulting in enhanced recognition and registration for three-dimensional surface matching.

MSCS-ICP: point cloud registration method using multi-view spatial coordinate system–ICP

“…However, the accuracy of this method largely depends on the extraction accuracy of edge points. Li et al [8] developed a deep learning-based key-point matching algorithm, strengthening recognition and alignment for 3D surface matching. But sometimes mismatched key point pairs are obtained, which leads to unsatisfactory results of 3D surface matching.…”

Low-overlap point cloud registration algorithm based on variable-length least trimmed squares

“…The ICP algorithm searches for matching point pairs using the nearest neighbor principle and calculates the pose transformation matrix based on the matching point pairs. Although ICP has high registration accuracy, it also has shortcomings such as high initial pose requirements, easy falling into local optima, and long calculation time [5] .…”

KdO-Net: Towards Improving the Efficiency of Deep Convolutional Neural Networks Applied in the 3D Pairwise Point Feature Matching

“…Recent works on the learning-based 3D point cloud matching have primarily focused on the 3D feature descriptors, including the learned 3D global feature descriptors [13,14], learned 3D local feature descriptors [9,12,[15][16][17][18][19][20][21], and weakly supervised feature descriptors [20,22]. For example, 3DMatch [12], one of the pioneer works with respect to the learning of 3D local descriptors, has been converted from its original point cloud into a volumetric 30 * 30 * 30 voxel grid of the Truncated Distance Function (TDF) values.…”

“…Although the learning-based 3D feature descriptors have achieved great success, the network architecture that is used for the feature extraction has not been the focus in previous studies. For example, Li et al [20] proposed a Multi-Statistics Histogram Descriptor (MSHD) that combines normal, curvature, and distribution density attribute features, but the extraction of the corresponding key points is only performed using a BP network. Owing to the difficulty in the acquisition of the ground-truth data and the application of deep learning in point cloud registration, besides the lack of a detailed presentation of the network structures and training strategies, higher performance methods for the application of the deep learning to the alignment are still being explored.…”