摘要
深度学习算法目前被广泛应用于激光雷达数据的分类研究中,但在处理过程中存在网络框架结构复杂、训练参数众多等问题。针对以上问题,本文对用于三维点云分析的无参数网络(Point-NN)进行了优化调整,将分类增强算法(CatBoost)融入其中,这样既可以利用Point-NN从原始激光点云中提取有效特征的优势,还可以利用CatBoost的优秀分类性能和泛化能力,从而显著提高点云特征的提取和聚合效率,简化特征学习网络的复杂度,同时提高点云的分类和识别精度。实验结果表明,在ModelNet40数据集上与Point-NN相比分类结果精度提升了2.0%,在ScanObjectNN中OBJ-BG、OBJ-ONLY和PB-T50-RS三个数据集上与原始组合相比分类精度分别提升了5.7%、4.8%和11.9%。
Deep learning algorithms are widely applied in LiDAR data classification research.However,there are issues such as complex network framework structures and numerous training parameters.To address these problems,this paper opti-mized the parameter-free network(Point-NN)used for three-dimensional(3D)point cloud analysis,integrating the classifi-cation boosting algorithm(CatBoost).This approach took advantage of Point-NN's ability to extract effective features from raw LiDAR point clouds and CatBoost's excellent classification performance and generalization capability.As a result,the method significantly improved the feature extraction and aggregation efficiency of point clouds,simplified the complexity of feature learning networks,and enhanced classification and recognition accuracy.Experimental results show that,compared to Point-NN,classification accuracy improved by 2.0%on the ModelNet40 dataset,and classification accuracy improved by 5.7%,4.8%,and 11.9%on the ScanObjectNN datasets(OBJ-BG,OBJ-ONLY,and PB-T50-RS)compared to the origi-nal combination.
作者
李林林
牛雁飞
杜跃飞
LI Linin;NIU Yanfei;DU Yuefei(Henan College of Surveying and Mapping,Zhengzhou,Henan 451464,China;Songshan Laboratory,Zhengzhou,Henan 450046,China)
出处
《北京测绘》
2025年第8期1159-1163,共5页
Beijing Surveying and Mapping
基金
河南省重大科技专项(221100211000-5)。
关键词
点云分类
无参数网络框架
特征提取
point cloud classification
parameter-free network framework
feature extraction
