Improved HDBSCAN-based cloud segmentation algorithm for large field attractions

Abstract Point cloud segmentation, as a core technology in 3D data processing, is essential in applications such as autonomous driving, smart cities, and industrial inspection. However, large-scale point cloud segmentation tasks often encounter boundary ambiguities among object categories. To address these challenges, this study introduces a large-scale segmentation framework combining local geometric features with an improved Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) algorithm. Initially, precise preprocessing of far-field noise is conducted using the spatial distribution characteristics of the 3D point cloud. Subsequently, Cloth Simulation Filtering (CSF) is utilized to extract initial ground points, further optimized by leveraging local ensemble features. Finally, non-ground points are pre-clustered through the enhanced HDBSCAN algorithm, which constructs a radius adjacency graph, quantifies spatial correlation via neighborhood overlap frequency, and incorporates a dual dynamic constraint mechanism based on connection weight thresholds and centroid distance thresholds. Experimental results demonstrate that integrating the geometric feature module significantly improves the ground segmentation method, increasing the F1 score from 86.73% to 97.34%. Furthermore, continuous multi-frame testing yielded an average F1 score of 94.31%. Notably, the improved HDBSCAN algorithm reduced the total number of clusters by 42.7% compared to the original method. By effectively combining geometric information with density-based clustering, the proposed framework significantly enhances segmentation accuracy and provides a robust and efficient solution for point cloud segmentation tasks in complex environments.