An Accurate and Robust Matching Network for Full/Partial Point Cloud Registration

In this work, we propose an Accurate and Robust Matching Network (ARM-Net) to facilitate the point cloud registration through accurate and robust point matching. The ARM-Net adopts two drastically different design concepts concerning network architectures. For one thing, ARM-Net utilizes dynamic graph convolution to extract features from evolving topology around a target point, where the neighboring information of that point is continuously changing. For another thing, ARM-Net uses the multi-layer perceptron (MLP) architecture to acquire "global" encoding of each local topology of the same point whose local topology remains constant. After extracting the features, the back-to-back transformers are used to identify self-attention and cross-attention between the two point clouds to be registered, and the matching subnetwork is then used to identify the points that correspond. Moreover, to obtain more local information and improve the distinctness of the extracted features, we propose a simple but effective module of local aggregation based on the K-Nearest Neighbor (KNN) algorithm before the extraction of the features. Providing noise-independent local information and improving registration performance is possible when K is sufficiently large for the local aggregation. For rotation and translation errors, simulation results demonstrate that ARM-Net outperforms other networks in terms of the MSE, RMSE and MAE. For rotation errors, ARM-Net has Root Mean Square Errors (RMSEs) of 0.0639 and 2.08962 for full and partial registration, respectively. For translation errors, ARM-Net has RMSEs of 2 x 10(-6) and 0.025574 for full and partial registration, respectively.