Image Guided Lidar Point Cloud Completion Algorithm

To address the issues of severe information loss and suboptimal fusion effects in multimodal feature extraction and integration during point cloud shape completion using autoen-coder structures, which results in difficulty in balancing local and global feature information of point clouds and significant loss of structural information in images, this paper proposes a LiDAR point cloud multi-scale completion algorithm guided by image rotation attention mechanisms, with a focus on the study of feature extraction from point clouds and images. The network employs an encoder-decoder structure, where the image feature extractor in the encoder utilizes rotation attention mechanisms to enhance the capability of image feature extraction. The point cloud feature extractor employs multi-scale methods to improve the global and local information of point cloud features and employs multi-level self-attention mechanisms to achieve multimodal feature fusion. The decoder then employs a multi-branch completion method to accomplish the point cloud completion task, with the network trained using chamfer distance guidance. Comparatively, our algorithm outperforms eight related algorithms on the ShapeNet-ViPC dataset across various metrics. Compared to the state-of-the-art network XMFnet, the category-averaged Chamfer Distance (CD) value is reduced by 11.71%. The proposed algorithm in this paper can better extract image structural information, and the feature extraction of incomplete point clouds can consider both global and local information. Furthermore, through multi-level progressive feature fusion, the algorithm enhances the complementarity of information between different modalities, leading to more accurate point cloud completion results.