A general and flexible point cloud simplification method based on feature fusion

Abstract

Large-scale, high-density point cloud data often pose challenges for direct application in various downstream tasks. To address this issue, this paper introduces a flexible point cloud simplification method based on feature fusion. After conducting a comprehensive analysis of the input point cloud, the method fuses the density feature that reflects point cloud uniformity with local geometric features that capture shape details. Based on the simplification objectives and fused feature values, the method optimizes the point distribution from a global perspective. Subsequently, by removing distance factors, purely local geometric features are incorporated into the farthest point sampling process and a feature-weighted voxel farthest point sampling algorithm is proposed to prioritize the preservation of local feature points. With a refined mechanism for adjusting point numbers, the method finally achieves fast and reasonable simplification of massive point clouds. Furthermore, extensive experiments have been designed to explore the impact of the features involved and their sensitivity to simplification results, offering detailed recommendations for parameter configuration. This method supports flexible transitions between global uniformity and heavy local feature preservation. Comparative results with previous studies demonstrate its excellent balance, exhibiting strong competitiveness in both output point cloud quality and computational efficiency. The core source code is publicly available at: https://github.com/chaojiale/PointCloudSimplification.