Position-aware feature matching algorithm based non-rigid point cloud registration
摘要
Existing non-rigid point cloud registration algorithms often encounter matching errors arise when dealing with similar regions in point clouds, and it is difficult to promptly filter out these misaligned corresponding points. To address these issues, a non-rigid point cloud registration algorithm based on position-aware feature matching is proposed. Firstly, the relative positions of the point cloud are encoded using Fourier transform, decomposing the 3D point cloud of non-rigid objects into feature space information and 3D positional information. This enables non-rigid point cloud registration for similar parts while preventing the loss of positional information during network iterations. Secondly, global information is summarized through the self-attention layers of the transformation blocks, and point cloud information exchange is facilitated through cross-attention layers to promote feature matching between the source and target point clouds. Next, we design and integrate an outlier removal strategy into a high-dimensional convolutional neural network to eliminate incorrect matching correspondences. The Welsch function is applied in the regularization term of the loss function to enhance the algorithm’s robustness against noise and partially overlapping point clouds. Finally, comparative experiments with seven existing algorithms on the 4Dmatch/4Dlomatch dataset demonstrate that our proposed method outperforms the second-best algorithm by 3.3