作者：高瑞贞1, 王诗浩1, 王皓乾1, 张京军1, 李志杰2
作者单位：1. 河北工程大学机械与装备工程学院，河北 邯郸 056038;
2. 河北新兴铸管股份有限公司，河北 邯郸 056300
关键词：图像处理;深度学习;图注意力机制;邻域点
摘要：
基于pointnet 网络的三维点云感知方法能够通过提取目标的几何特征信息来对目标进行分类。虽然pointnet 网络能够提取到点云数据的局部特征，但未考虑点与其邻域点之间的关系，因此一旦缺失一个点的局部特征，网络的性能会受到较大的影响。针对这一问题，该文提出一种基于图注意力机制(graphic attention mecahnism)的新网络架构ga-pointnet 。模型利用图注意力机制在点与其邻域点之间分配注意力系数，完成点云局部特征的提取。在分类实验中，该文在modelnet40数据集上的实验结果表明，提出的ga-pointnet 模型最终的平均分类准确率达到了88.8%，总体准确率达到了91.3%。相较于pointnet 基线模型总体准确率提高1.1百分点，验证了ga-pointnet 在分类任务中的有效性。
3d point cloud perception based on graph attention mechanism
gao ruizhen1, wang shihao1, wang haoqian1, zhang jingjun1, li zhijie2
1. college of mechanical and equipment engineering, hebei university of engineering, handan 056038, china;
2. hebei xinxing casting pipe company limited, handan 056300, china
abstract: the 3d point cloud perception method based on pointnet network can classify objects by extracting geometric feature information of objects. although the pointnet network can extract the local features of point cloud data, the relationship between points and their neighborhood points is not considered. thus, once the local feature of a point is missing, the performance of the network will be greatly affected. aming at solving this problem, a new network architecture based on graphic attention mechanism(ga-pointnet ) is proposed. the model uses the graph attention mechanism to distribute the attention coefficient between points and their neighborhoods points, and extracts the local features finally. in the classification experiment, results on modelnet40 data set show that the mean class accuracy of the model is 88.8%, and the overall accuracy rate reached 91.3%, improved by 1.1% compare with pointnet baseline model. the results confirm the effectiveness of the proposed ga-pointnet model in classification tasks.
keywords: image processing;deep learning;graph attention mechanism;neighborhood points
2024, 50(7):155-162 收稿日期: 2022-06-25;收到修改稿日期: 2022-08-19
基金项目: 河北省高校科技攻关项目(zd2018207)
作者简介: 高瑞贞（1979-），男，河北邯郸市人，博士，教授，博士生导师，研究方向为深度学习。
参考文献
[1] 田永林, 沈宇, 李强, 等. 平行点云: 虚实互动的点云生成与三维模型进化方法[j]. 自动化学报, 2020, 46(12): 2572-2582.
tian y l, shen y, li q, et al. parallel point clouds:point clouds generation and 3d model evolution via virtual-real interaction[j]. acta automatica sinica, 2020, 46(12): 2572-2582.
[2] 张越, 翟福琪, 蔡孙宝, 等. 基于点云数据的植物叶片特征提取及三维重建[j]. 中国测试, 2021, 47(8): 6-12.
zhang y, zhao f q, cai s b, et al. feature extraction and 3d reconstruction of plant leaf based on point cloud data[j]. china measurement & test, 2021, 47(8): 6-12.
[3] 廖瑞杰, 杨绍发, 孟文霞, 等. seggraph: 室外场景三维点云闭环检测算法[j]. 计算机研究与发展, 2019, 56(2): 338-348.
liao r j, yang s w, meng w x, et al. seggraph: an algorithm for loop-closure detection in outdoor scenes using 3d point clouds[j]. journal of computer research and development, 2019, 56(2): 338-348.
[4] 庄仁诚, 陈鹏, 傅瑶, 等. 列车车轮三维结构光检测中的点云处理研究[j]. 中国测试, 2021, 47(2): 19-25.
zhuang r c, chen p, fu y, et al. research on point cloud processing in train wheels three-dimensional structured light inspection[j]. china measurement & test, 2021, 47(2): 19-25.
[5] bold n, zhang c, akashi t. 3d point cloud retrieval with bidirectional feature match[j]. ieee access, 2019, 7: 164194-164202.
[6] prakhya s m, liu b, lin w, et al. b-shot: a binary 3d feature descriptor for fast keypoint matching on 3d point clouds[j]. autonomous robots, 2017, 41(7): 1501-1520.
[7] 鲁斌, 范晓明. 基于改进自适应k均值聚类的三维点云骨架提取的研究[j]. 自动化学报, 2022, 48(8): 1994-2006.
lu b, fan x m. research on 3d point cloud skeleton extraction based on improved adaptive k-means clustering[j]. acta automatica sinica, 2022, 48(8): 1994-2006.
[8] 高继东, 焦鑫, 刘全周, 等. 机器视觉与毫米波雷达信息融合的车辆检测技术[j]. 中国测试, 2021, 47(10): 33-40.
gao j, jiao x, liu q, et al. research on vehicle detection based on data fusion of machine vision and millimeter wave radar[j]. china measurement & test, 2021, 47(10): 33-40.
[9] 王昕, 赵飞, 蒋佐富, 等. 迁移学习和卷积神经网络电力设备图像识别方法[j]. 中国测试, 2020, 46(5): 108-113.
wang x, zhao f, jiang z, et al. power equipment image recognition method based on transfer learning and convolutional neural network[j]. china measurement & test, 2020, 46(5): 108-113.
[10] 彭熙舜, 陆安江, 唐鑫鑫, 等. 三维激光点云下利用mean_shift的欧式目标分割[j]. 激光杂志, 2022, 43(2): 119-123.
peng x s, lu a j, tang x x, et al. euclidean target segmentation using mean_shift under 3d laser point cloud[j]. laser journal, 2022, 43(2): 119-123.
[11] lecun y, boser b, denker j s, et al. backpropagation applied to handwritten zip code recognition[j]. neural computation, 1989, 1(4): 541-551.
[12] nurunnabi a, teferle f n, li j, et al. investigation of pointnet for semantic segmentation of large-scale outdoor point clouds[j]. isprs-international archives of the photogrammetry, remote sensing and spatial information sciences, 2021, 46: 397-404.
[13] liu r, ren l, wang f. 3d point cloud of single tree branches and leaves semantic segmentation based on modified pointnet network[c]//journal of physics: conference series. iop publishing, 2021, 2074(1): 012026.
[14] 白静, 司庆龙, 秦飞巍. 轻量级实时点云分类网络lightpointnet[j]. 计算机辅助设计与图形学学报, 2019, 31(4): 612-621.
bai j, si q l, qin f w. lightweight real-time point cloud classification network lightpointnet[j]. journal of computer-aided design & computer graphics, 2019, 31(4): 612-621.
[15] qi c r, yi l, su h, et al. pointnet deep hierarchical feature learning on point sets in a metric space[c]//proceedings of the 31st international conference on neural information processing systems, 2017.
[16] shin y h, son k w, lee d c. semantic segmentation and building extraction from airborne lidar data with multiple return using pointnet [j]. applied sciences, 2022, 12(4): 1975.
[17] ravanbakhsh s, schneider j, poczos b. deep learning with sets and point clouds[j/ol]. arxiv preprint arxiv: 1611.04500, 2016. [2022-06-02]. https://arxiv.org/abs/ 1611.04500
[18] su h, jampani v, sun d, et al. splatnet: sparse lattice networks for point cloud processing[c]//proceedings of the ieee conference on computer vision and pattern recognition, 2018.
[19] hua b s, tran m k, yeung s k. pointwise convolutional neural networks[c]//proceedings of the ieee conference on computer vision and pattern recognition. 2018: 984-993.
[20] chen y, liu g, xu y, et al. pointnet network architecture with individual point level and global features on centroid for als point cloud classification[j]. remote sensing, 2021, 13(3): 472.
[21] zhang l, wang h. a novel segmentation method for cervical vertebrae based on pointnet and converge segmentation[j]. computer methods and programs in biomedicine, 2021, 200: 105798.
[22] wu z, song s, khosla a, et al. 3d shapenets: a deep representation for volumetric shapes[c]//proceedings of the ieee conference on computer vision and pattern recognition, 2015.
