Multi-Level Feature Extraction Point Cloud Registration Model Based on Cross-Attention Matrix

Three-dimensional point cloud technology is essential for intelligent robots, autonomous driving, and human-computer interaction. Due to the self-occlusion issue, a single scan cannot acquire complete point cloud data, requiring the integration of point cloud data scanned from different directions through rotational and translational transformations. For partial overlap point cloud registration, not every point in the source point cloud can find a corresponding relationship in the target point cloud, posing significant challenges and complexity to point cloud registration. This paper proposes a point cloud registration model based on multi-level feature extraction using a cross-attention matrix to address the issue of partial overlap in point cloud registration. The cross-attention matrix is introduced within the feature extraction network to mix contextual information between layers, seeking correspondences in the overlapping areas and optimizing the construction of matching features. To enhance the internal dependencies within the point cloud and maintain internal associations, a Transformer encoder is proposed to follow the graph convolution in each layer, and a complete Transformer is added after the final feature extraction to integrate overall feature contextual information, which can suppress the impact of noise on the registration effect. Experimental results show that on the ModelNet40 dataset, compared to other network models, the proposed model achieves higher precision in point cloud registration.