基于轻量级网络和加权RF的点云分类分割算法

doi:10.11996/JG.j.2095-302X.2026010143

摘要/Abstract

摘要：

针对点云分类分割方法中存在的高计算开销、复杂网络模型等问题，提出一种基于轻量级网络和加权随机森林(RF)的点云分类分割算法。该算法采用层次化的方式实现高效分类分割，首先针对传统神经网络层数多、计算复杂等问题，构造轻量级神经网络，并利用其提取点云的全局形状、区域间关系、曲率、法向量和颜色等特征，实现点云的快速粗分类分割；然后针对数据不平衡的问题，设计自适应分类分割策略，并引入加权RF，结合不一致度量筛选与动态加权优化机制，以实现点云精分类分割。在ModelNet40数据集上进行分类实验，在Semantic3D数据集和室外场景点云数据上进行了分割实验，结果表明，相比Local Geo-Transformer，PointNeXt和FastPointNet++等算法，该算法的分类分割精度分别提高了约1.9%，1.6%和1.7%，分类分割时间分别降低了约40%，30%和20%。由此可见，基于轻量级网络和加权RF的点云分类分割算法在保持较高分类分割精度的同时，可以有效缩短模型的训练时间，提高分类分割效率，是一种有效的点云分类分割算法。

关键词: 点云分类分割, 轻量级网络, 加权随机森林, 全局形状, 曲率

Abstract:

To address the issues of high computational cost and complex network models in point cloud classification and segmentation methods, a point cloud classification and segmentation algorithm based on lightweight networks and weighted Random Forest (RF) was proposed. The algorithm achieved efficient classification and segmentation in a hierarchical manner. Firstly, to address the issues of multiple layers and complex computation in traditional neural networks, a lightweight neural network was constructed to extract point cloud features such as global shape, inter-regional relationships, curvature, normal vector, and color, thereby achieving rapid rough classification and segmentation of point clouds. Then, to address data imbalance, an adaptive classification and segmentation strategy was designed. By introducing a weighted RF and combining inconsistency-measurement screening with dynamic-weighting optimization mechanisms, fine classification and segmentation of point clouds were achieved. The algorithm conducted classification experiments on the ModelNet40 dataset and segmentation experiments on the Semantic3D dataset and outdoor-scene point-cloud data. The results showed that compared with Local Geo-Transformer, PointNeXt, and FastPointNet++, classification and segmentation accuracy increased by approximately 1.9%, 1.6%, and 1.7%, respectively, while classification and segmentation time was reduced by approximately 40%, 30%, and 20%, respectively. Thus, the proposed point-cloud classification and segmentation algorithm based on lightweight networks and weighted RF can effectively reduce the training time of the model and improve the efficiency of classification and segmentation while maintaining high accuracy, making it an effective point cloud classification and segmentation algorithm.

Key words: point cloud classification and segmentation, lightweight network, weighted random forest, global shape, curvature

中图分类号:

赵夫群, 郝寒竹, 余佳乐. 基于轻量级网络和加权RF的点云分类分割算法[J]. 图学学报, 2026, 47(1): 143-151.

ZHAO Fuqun, HAO Hanzhu, YU Jiale. A point cloud classification and segmentation algorithm based on lightweight networks and weighted RF[J]. Journal of Graphics, 2026, 47(1): 143-151.

图/表 10

参考文献 24

[1]	贾明超, 冯斌, 吴鹏, 等. 一种融合改进A*算法与改进动态窗口法的文旅服务机器人路径规划[J]. 图学学报, 2024, 45(3): 505-515.
	JIA M C, FENG B, WU P, et al. A path planning for cultural tourism service robot combining improved A* algorithm and improved dynamic window approach[J]. Journal of Graphics, 2024, 45(3): 505-515 (in Chinese). DOI
[2]	杨振, 李俊丽, 杨立炜, 等. 安全性A*融合DWA的分布式多移动机器人路径规划方法[J]. 控制工程, 2024, 31(12): 2284-2295.
	YANG Z, LI J L, YANG L W, et al. Distributed multi-mobile robots path planning method based on safe A* fused with dynamic window approach[J]. Control Engineering of China, 2024, 31(12): 2284-2295 (in Chinese).
[3]	成欢, 王硕, 李孟, 等. 面向自动驾驶场景的神经辐射场综述[J]. 图学学报, 2023, 44(6): 1091-1103. DOI
	CHENG H, WANG S, LI M, et al. A review of neural radiance field for autonomous driving scene[J]. Journal of Graphics, 2023, 44(6): 1091-1103 (in Chinese). DOI
[4]	MENG X H, LU X Y, YE H L, et al. A new self-augment CNN for 3D point cloud classification and segmentation[J]. International Journal of Machine Learning and Cybernetics, 2024, 15(3): 807-818. DOI
[5]	WEN J J, MA J, ZHAO Y H, et al. Point cloud semantic segmentation based on local feature fusion and multilayer attention network[J]. IET Computer Vision, 2024, 18(3): 381-392. DOI URL
[6]	ZENG Z Y, XU Y Y, XIE Z, et al. RG-GCN: a random graph based on graph convolution network for point cloud semantic segmentation[J]. Remote Sensing, 2022, 14(16): 4055. DOI URL
[7]	LIU H, TIAN S H. Deep 3D point cloud classification and segmentation network based on GateNet[J]. The Visual Computer, 2024, 40(2): 971-981. DOI
[8]	LI S, WANG M, WU J, et al. CloudDeepLabV3+: a lightweight ground-based cloud segmentation method based on multi-scale feature aggregation and multi-level attention feature enhancement[J]. International Journal of Remote Sensing, 2023, 44(15): 4836-4856. DOI URL
[9]	JING Y X, SHEIL B, ACIKGOZ S. A lightweight transformer-based neural network for large-scale masonry arch bridge point cloud segmentation[J]. Computer-Aided Civil and Infrastructure Engineering, 2024, 39(16): 2427-2438. DOI URL
[10]	QIAN G C, LI Y C, PENG H W, et al. PointNeXt: revisiting PointNet++ with improved training and scaling strategies[C]// The 36th International Conference on Neural Information Processing Systems. New York: Curran Associates Inc., 2022: 23192-23204.
[11]	CHEN X, LIU K, ZHANG Z. A PointNet-based CFAR detection method for radar target detection in sea clutter[J]. IEEE Geoscience and Remote Sensing Letters, 2024, 21: 3502305.
[12]	ALFIO V S, PEPE M, COSTANTINO D. The use of random forest for the classification of point cloud in urban scene[J]. International Journal of Engineering Trends and Technology, 2024, 72(3): 1-9.
[13]	ZHAO F Q, HUANG H, XIAO N N, et al. A point cloud segmentation algorithm based on multi-feature training and weighted random forest[J]. Measurement Science and Technology, 2025, 36(1): 015407. DOI
[14]	MOHAMED M, MORSY S, EL-SHAZLY A. Improvement of 3D LiDAR point cloud classification of urban road environment based on random forest classifier[J]. Geocarto International, 2022, 37(27): 15604-15626. DOI URL
[15]	FENG G Q, LI W L, ZHAO X L, et al. LessNet: lightweight and efficient semantic segmentation for large-scale point clouds[J]. IET Cyber-Systems and Robotics, 2022, 4(2): 107-115. DOI URL
[16]	ZHAO X, WANG L M, ZHANG Y F, et al. A review of convolutional neural networks in computer vision[J]. Artificial Intelligence Review, 2024, 57(4): 99. DOI
[17]	RÖSSIG A, PETKOVIC M. Advances in verification of ReLU neural networks[J]. Journal of Global Optimization, 2021, 81(1): 109-152. DOI
[18]	JALAL N, MEHMOOD A, CHOI G S, et al. A novel improved random forest for text classification using feature ranking and optimal number of trees[J]. Journal of King Saud University-Computer and Information Sciences, 2022, 34(6): 2733-2742. DOI URL
[19]	吴冬, 阎卫东, 王井利. 基于特征重要性加权的随机森林点云分类研究[J]. 电子测量技术, 2023, 46(20): 120-127.
	WU D, YAN W D, WANG J L. Random forest point cloud classification algorithm based on feature importance weighting[J]. Electronic Measurement Technology, 2023, 46(20): 120-127 (in Chinese).
[20]	LIANG G, ZHAO X L, ZHAO J H, et al. MVCNN: a deep learning-based ocean-land waveform classification network for single-wavelength LiDAR bathymetry[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023, 16: 656-674. DOI URL
[21]	MOHAMMADI F G, CHEN C, SHENAVARMASOULEH F, et al. 3D-model shapenet core classification using meta-semantic learning[EB/OL]. [2025-01-10]. https://arxiv.org/abs/2205.15869.
[22]	MSALLAM M, SYRYAMKIN V I. Comparison of classical and deep learning classifiers for 3D objects recognition[EB/OL]. [2025-01-10]. https://elibrary.ru/wgdizx.
[23]	SHCHENYAVSKAYA L A, GURA D A, DYACHENKO R A. Analysis of the PointNet neural network architecture[J]. Herald of Dagestan State Technical University Technical Sciences, 2024, 50(4): 158-165. DOI URL
[24]	HELL M, BRANDMEIER M, BRIECHLE S, et al. Classification of tree species and standing dead trees with lidar point clouds using two deep neural networks: pointCNN and 3DmFV-net[J]. PFG-Journal of Photogrammetry, Remote Sensing and Geoinformation Science, 2022, 90(2): 103-121. DOI

模型	OA/%	F1	IR/%
原始RF	85.30	83.9	─
静态特征权重优化RF	87.10	85.5	15.6
静态分类分割错误率优化RF	86.80	85.0	10.3
动态联合评估权重优化RF	89.96	88.1	32.8

模型	OA/%	F1	IR/%
原始RF	85.30	83.9	─
静态特征权重优化RF	87.10	85.5	15.6
静态分类分割错误率优化RF	86.80	85.0	10.3
动态联合评估权重优化RF	89.96	88.1	32.8

数据集	网络	参数量	OA/%	训练时间/min	单个物体识别时间/ms
ModelNet40	MVCNN^[20]	1.280 0×10⁸	91.10	500~800	8.40
	3DshapeNets^[21]	3.400 0×10⁷	78.52	1200	15.00
	VoxNet^[22]	8.900 0×10⁵	84.02	360~720	5.20
	PointNet^[23]	3.540 0×10⁶	88.53	254	0.82
	PointCNN^[24]	7.850 0×10⁵	89.32	600~800	1.50
	PointNet++^[7]	1.480 0×10⁶	90.50	663	11.23
	RG-GCN^[6]	2.200 0×10⁶	91.00	700~900	2.50
	Local Geo-Transformer^[9]	2.500 0×10⁵	92.50	193	0.54
	PointNeXt^[10]	4.200 0×10⁵	92.81	180	0.46
	FastPointNet++^[11]	2.500 0×10⁵	91.57	150	0.42
	本文算法	2.359 4×10⁵	93.23	143	0.39
Semantic3D	MVCNN^[20]	1.150 0×10⁸	83.80	550~750	7.20
	3DshapeNets^[22]	3.500 0×10⁷	77.50	1100	17.00
	VoxNet^[22]	1.200 0×10⁶	81.80	270	5.40
	PointNet^[23]	4.200 0×10⁶	86.50	200	3.00
	PointCNN^[24]	8.300 0×10⁵	88.50	375~525	5.80
	PointNet++^[7]	1.450 0×10⁶	89.90	640	8.30
	RG-GCN^[6]	1.400 0×10⁶	88.80	550~750	6.80
	Local Geo-Transformer^[9]	5.200 0×10⁵	90.10	90	1.40
	PointNeXt^[10]	6.500 0×10⁵	90.51	80	1.26
	FastPointNet++^[11]	3.800 0×10⁵	89.83	65	0.90
	本文算法	2.500 0×10⁵	91.80	55	0.70
室外场景点云数据	MVCNN^[20]	1.100 0×10¹¹	84.56	1700~2400	11.40
	3DshapeNets^[22]	3.500 0×10⁹	78.21	2900	21.20
	VoxNet^[22]	1.500 0×10⁹	79.34	900	7.56
	PointNet^[23]	3.000 0×10⁸	84.62	820	5.60
	PointCNN^[24]	8.520 0×10⁷	87.24	950~1200	8.12
	PointNet++^[7]	1.200 0×10⁹	88.27	1450	10.34
	RG-GCN^[6]	2.100 0×10⁸	88.43	1400~1600	8.23
	Local Geo-Transformer^[9]	3.318 3×10⁶	88.96	190	3.40
	PointNeXt^[10]	3.200 0×10⁶	89.11	145	2.30
	FastPointNet++^[11]	3.200 0×10⁶	89.56	164	2.90
	本文算法	3.012 3×10⁶	92.24	90	1.46

数据集	网络	参数量	OA/%	训练时间/min	单个物体识别时间/ms
ModelNet40	MVCNN^[20]	1.280 0×10⁸	91.10	500~800	8.40
	3DshapeNets^[21]	3.400 0×10⁷	78.52	1200	15.00
	VoxNet^[22]	8.900 0×10⁵	84.02	360~720	5.20
	PointNet^[23]	3.540 0×10⁶	88.53	254	0.82
	PointCNN^[24]	7.850 0×10⁵	89.32	600~800	1.50
	PointNet++^[7]	1.480 0×10⁶	90.50	663	11.23
	RG-GCN^[6]	2.200 0×10⁶	91.00	700~900	2.50
	Local Geo-Transformer^[9]	2.500 0×10⁵	92.50	193	0.54
	PointNeXt^[10]	4.200 0×10⁵	92.81	180	0.46
	FastPointNet++^[11]	2.500 0×10⁵	91.57	150	0.42
	本文算法	2.359 4×10⁵	93.23	143	0.39
Semantic3D	MVCNN^[20]	1.150 0×10⁸	83.80	550~750	7.20
	3DshapeNets^[22]	3.500 0×10⁷	77.50	1100	17.00
	VoxNet^[22]	1.200 0×10⁶	81.80	270	5.40
	PointNet^[23]	4.200 0×10⁶	86.50	200	3.00
	PointCNN^[24]	8.300 0×10⁵	88.50	375~525	5.80
	PointNet++^[7]	1.450 0×10⁶	89.90	640	8.30
	RG-GCN^[6]	1.400 0×10⁶	88.80	550~750	6.80
	Local Geo-Transformer^[9]	5.200 0×10⁵	90.10	90	1.40
	PointNeXt^[10]	6.500 0×10⁵	90.51	80	1.26
	FastPointNet++^[11]	3.800 0×10⁵	89.83	65	0.90
	本文算法	2.500 0×10⁵	91.80	55	0.70
室外场景点云数据	MVCNN^[20]	1.100 0×10¹¹	84.56	1700~2400	11.40
	3DshapeNets^[22]	3.500 0×10⁹	78.21	2900	21.20
	VoxNet^[22]	1.500 0×10⁹	79.34	900	7.56
	PointNet^[23]	3.000 0×10⁸	84.62	820	5.60
	PointCNN^[24]	8.520 0×10⁷	87.24	950~1200	8.12
	PointNet++^[7]	1.200 0×10⁹	88.27	1450	10.34
	RG-GCN^[6]	2.100 0×10⁸	88.43	1400~1600	8.23
	Local Geo-Transformer^[9]	3.318 3×10⁶	88.96	190	3.40
	PointNeXt^[10]	3.200 0×10⁶	89.11	145	2.30
	FastPointNet++^[11]	3.200 0×10⁶	89.56	164	2.90
	本文算法	3.012 3×10⁶	92.24	90	1.46