A feature-preserving simplification method for dense point clouds based on voxel priority filtering

Abstract The inherently high-dimensional and information-dense nature of point clouds requires substantial storage and computing resources, which severely limits their widespread applicability. It is crucial to simplify the point cloud data efficiently and accurately for various downstream tasks. To address the challenges of detail loss and time consumption, we propose a novel method that utilizes voxel filtering and partition sampling strategy for point cloud simplification. The voxel grid structure of original point cloud is established based on the spatial location of the points. The voxels are then classified as feature and non-feature voxels by calculating the Eigenentropy on the Gaussian sphere. Combined with an effective partition strategy for further sampling, the proposed method ensures global geometric consistency and preserves distinct features in the simplified point cloud. To evaluate the performance comprehensively, we conducted extensive experiments on the models of standard, industrial real-world scanning and natural scene datasets. Compared with several relevant approaches, the results demonstrate that the proposed method significantly enhances the flexibility and accuracy of point cloud simplification. It exhibits superior capabilities in engineering applications such as visualization, 3D reconstruction, and point cloud registration.