Vertex Convolution-Based Planar-Dominant Point Cloud Segmentation Algorithm

With the rapid advancement of 3D perception technologies, point cloud segmentation has emerged as a pivotal task in critical applications spanning robotic environmental awareness, autonomous driving scene interpretation, and digital twin construction. While traditional approaches primarily rely on manually engineered heuristic rules, rendering them susceptible to noise interference and inadequate for handling intricate geometric configurations, contemporary deep learning paradigms such as PointNet and PointNet++ have demonstrated remarkable progress through end-to-end feature extraction architectures. Nevertheless, these data-driven methods remain fundamentally constrained to category-specific semantic segmentation scenarios. To bridge this methodological gap, we present a novel class-agnostic segmentation framework that synergistically integrates a learnable region-growing mechanism with geometric consistency regularization. Our approach introduces three principal innovations: dynamic region growing modeled using a dualchannel encoder-decoder network, a geometric consistency loss function that enhances boundary accuracy by incorporating normal smoothness and curvature continuity, and cross-domain generalization validated on the S3DIS and ScanNet datasets. Experimental results demonstrate that the proposed approach achieves high robustness and precision, particularly in planedominated scenes, significantly outperforming traditional regiongrowing algorithms and deep learning baselines.