Point Cloud Registration Based on AGConv Local Feature Descriptors

Abstract

To improve the registration accuracy of existing point cloud registration models for real point cloud data, an improved point cloud registration model is proposed based on the local feature descriptors of Adaptive Graph Convolution(AGConv). In the data preprocessing module, the sampling points in the local patch are normalized by constructing the sampling points in the point cloud and calculating the Local Reference Frame(LRF) to render them insensitive to rotation transformation. In the feature extraction module, AGConv is used to generate an adaptive kernel for the sampling points. The relationships between the points of different semantic parts are fully excavated. The standardized local patches are then input into the AGConv-based feature extraction network to calculate the local feature descriptors and improve the robustness of the local features to occlusion and clutter. In the point cloud registration module, the RANdom SAmpling Consistency(RANSAC) algorithm is used to estimate a rigid transformation matrix. The experimental results on the 3DMatch dataset show that, compared with the DIP model, the Feature Matching Recall(FMR) of this model is increased by 2.3 percentage points, and the Registration Recall(RR) is increased by 5 percentage points. This can effectively improve the registration accuracy of point clouds with good robustness.