基于邻域特征编码优化的液压支架激光点云分割算法  

Laser point cloud segmentation algorithm for hydraulic support based on neighborhood feature encoding and optimization

作　　者：王俊甫 薛晓杰 杨艺[2,3] WANG Junfu;XUE Xiaojie;YANG Yi(Zhengzhou Hengda Intelligent Control Technology Company Limited,Zhengzhou 450000,China;School of Electrical Engineering and Automation,Henan Polytechnic University,Jiaozuo 454003,China;Henan Key Laboratory of Intelligent Detection and Control of Coal Mine Equipment,Henan Polytechnic University,Jiaozuo 454003,China)

机构地区：[1]郑州恒达智控科技股份有限公司,河南郑州450000 [2]河南理工大学电气工程与自动化学院,河南焦作454003 [3]河南理工大学河南省煤矿装备智能检测与控制重点实验室,河南焦作454003

出　　处：《工矿自动化》2024年第7期98-106,178,共10页Journal Of Mine Automation

基　　金：河南省科技攻关项目(232102210040)。

摘　　要：受井下煤尘和易被遮挡的影响,液压支架激光点云数据容易出现残缺。现有点云分割算法难以获取细粒度的点云特征,无法得到完整的点云结构信息,且易在邻域内引入语义信息不相似的点,导致液压支架激光点云分割精度低。针对上述问题,提出了一种基于邻域特征编码优化的液压支架激光点云分割算法。引入了由邻域特征编码模块、邻域特征优化模块和混合池化模块组成的局部邻域特征聚合模块:邻域特征编码模块在传统三维坐标编码的基础上加入极坐标编码和质心偏移来表征局部点云空间结构,提升对残缺点云的特征提取能力;邻域特征优化模块通过特征距离判断并丢弃冗余特征,来优化邻域空间内的特征表达,从而更有效地学习点云局部细粒度特征,增强点云局部上下文信息;混合池化模块结合注意力池化和最大池化,通过聚合邻域内的显著特征和重要特征来获取具有丰富信息的单点特征,减少信息丢失。构建了由2组局部邻域特征聚合模块和残差连接组成的邻域扩张模块,以捕获特征间的长距离依赖关系,扩大单个点的局部感受野,并聚合更多有效特征。实验结果表明,该算法在液压支架激光点云分割数据集上的平均交并比为93.26%,平均准确率为96.42%,可有效区分液压支架不同的几何结构,实现液压支架各部件的准确分割。Due to the influence of underground coal dust and easy obstruction,the laser point cloud data of hydraulic supports is prone to be incomplete.The existing point cloud segmentation algorithms are difficult to obtain fine-grained point cloud features,unable to obtain complete structural information of the point cloud.The algorithms are prone to introducing semantically dissimilar points in the neighborhood,resulting in low precision of laser point cloud segmentation for hydraulic supports.In order to solve the above problems,a laser point cloud segmentation algorithm for hydraulic supports based on neighborhood feature encoding and optimization is proposed. The method introduces a local neighborhood feature aggregation module consisting of neighborhoodfeature encoding module, neighborhood feature optimization module, and hybrid pooling module. Theneighborhood feature encoding module adds polar coordinate encoding and centroid offset to represent the spatialstructure of local point clouds on the basis of traditional 3D coordinate encoding, improving the feature extractioncapability for incomplete point clouds. The neighborhood feature optimization module optimizes the featureexpression in the neighborhood space by judging the feature distance and discarding redundant features, therebymore effectively learning the local fine-grained features of the point cloud and enhancing the local contextualinformation of the point cloud. The hybrid pooling module combines attention pooling and max pooling to obtainsingle point features with rich information by aggregating salient and important features within the neighborhood,reducing information loss. A neighborhood expansion module consisting of two sets of local neighborhood featureaggregation modules and residual connections is constructed to capture long-range dependencies between features,expand the local receptive field of individual points, and aggregate more effective features. The experimentalresults show that the algorithm has an mean intersection over union of 93.26%

关 键 词：液压支架 激光点云 点云分割 邻域特征编码 邻域特征优化 

分 类 号：TD355[矿业工程—矿井建设]