A novel partial point cloud registration method based on graph attention network

Point cloud registration is a challenging task due to sparsity and unknown initial correspondence information. The traditional registration methods tend to converge to local optimal solutions and rely on good initial correspondence information. Deep learning-based methods show good adaptability to initial information and noises, but they cannot effectively cope with partial-to-partial registration scenes. This paper proposes a partial point cloud registration method based on graph attention network. The context information of the point cloud is obtained by a message passing mechanism. The attention features of the key registration points are extracted by an attention network. The key matching points are chosen by a key point selection module. Virtual correspondences are generated based on these key points and their features. A rigid transformation is obtained based on the virtual registration by a singular value decomposition layer. The performance of the proposed method is evaluated in three scenarios based on the ModelNet40 dataset. Experimental results show that the proposed method is robust to arbitrary initial positions and noises. It obtains higher registration accuracy than traditional methods while maintaining low network complexity.

Automated summary

This paper proposes a partial point cloud registration method based on a graph attention network. The context information of the point cloud is obtained through message passing and key point selection modules. Virtual correspondences are generated based on key points and their features, and a rigid transformation is obtained using a singular value decomposition layer. Experimental results demonstrate that the proposed method is robust to arbitrary initial positions and noises, and achieves higher registration accuracy while maintaining low network complexity.