Li Yong, Luo Yinzheng, Lian Dehang, Bu Chunning, Wang Hongxiang
Guangxi Key Laboratory of Intelligent Control and Maintenance of Power Equipment, School of Electrical Engineering, Guangxi University, Nanning, 530004, China.
Guangxi Key Laboratory of Forest Ecology and Conservation, Nanning, 530004, China.
Heliyon. 2025 Feb 11;11(4):e42623. doi: 10.1016/j.heliyon.2025.e42623. eCollection 2025 Feb 28.
Point cloud classification, as one of the key techniques for point cloud data processing, is an important step for the application of point cloud data. However, single-point-based point cloud classification faces the challenge of poor robustness, and single-scale point clusters only consider a single neighborhood, leading to insufficient feature representation. In addition, numerous cluster-based classification methods require further development in constructing point clusters and extracting their features for representing point cloud objects. To address these issues, we present a point cloud classification algorithm based on multi-level aggregated features. In our method, we employ a multi-level point cluster construction approach based on MLPCS (Multi-level Point Cluster Segmentation), which divides the original point cloud into three different levels of point clusters by voxel downsampling and rescanning, Voxel-Meanshift, and Voxel-DBSCAN (Density-Based Spatial Clustering of Applications with Noise). The features for each level of point clusters are extracted and their representation is improved by adopting methods like max pooling, Bag of Words (BoW) and K-Means. The multi-level point cluster features are then aggregated and combined with a random forest classifier to achieve automatic classification of point clouds. Finally, we conducted ablation and comparison experiments to verify the effectiveness and advantages of the algorithm. Our method achieved classification accuracy/Kappa coefficient of 99.88 %/99.86 % and 93.44 %/83.61 % respectively in the experiments on two sets of large-scale outdoor scene data, and the ablation experiments confirmed the effectiveness of our algorithm.
点云分类作为点云数据处理的关键技术之一，是点云数据应用的重要环节。然而，基于单点的点云分类面临着鲁棒性差的挑战，单尺度点簇仅考虑单个邻域，导致特征表示不足。此外，众多基于聚类的分类方法在构建点簇和提取其特征以表示点云对象方面需要进一步发展。为了解决这些问题，我们提出了一种基于多级聚合特征的点云分类算法。在我们的方法中，我们采用基于MLPCS（多级点簇分割）的多级点簇构建方法，通过体素下采样和重新扫描、体素均值漂移和体素DBSCAN（基于密度的带有噪声的空间聚类应用）将原始点云划分为三个不同级别的点簇。通过采用最大池化、词袋模型（BoW）和K均值等方法，提取每个级别的点簇特征并改进其表示。然后将多级点簇特征进行聚合，并与随机森林分类器相结合，实现点云的自动分类。最后，我们进行了消融实验和对比实验，以验证该算法的有效性和优势。在两组大规模户外场景数据的实验中，我们的方法分别实现了99.88%/99.86%和93.44%/83.61%的分类准确率/Kappa系数，消融实验证实了我们算法的有效性。
