基于膨胀图卷积与离群点过滤的残缺点云配准  

Partial Point Cloud Registration Based on Dilated Graph Convolution and Outlier Filtering

作　　者：孙战里 张玉欣 陈霞[2,4] SUN Zhanli;ZHANG Yuxin;CHEN Xia(School of Artificial Intelligence,Anhui University,Hefei 230601,China;Key Laboratory of Multi-Modal Cognitive Computing of Anhui Province,Anhui University,Hefei 230601,China;School of Electrical Engineering and Automation,Anhui University,Hefei 230601,China;School of Information and Computer,Anhui Agricultural University,Hefei 230036,China)

机构地区：[1]安徽大学人工智能学院,合肥230601 [2]安徽大学多模态认知计算安徽省重点实验室,合肥230601 [3]安徽大学电气工程与自动化学院,合肥230601 [4]安徽农业大学信息与计算机学院,合肥230036

出　　处：《计算机工程与应用》2022年第22期186-194,共9页Computer Engineering and Applications

基　　金：国家自然科学基金(61972002);多模态认知计算安徽省重点实验室(安徽大学)开放课题(MMC202004)。

摘　　要：由于点云在非欧几里德空间中,受到结构不规则、噪声、离群点等不利因素的影响,如何准确配准残缺点云,仍然是一个具有挑战性的任务。针对此任务,提出了一种有效的残缺点云配准网络。为了有效提取局部点云的细粒度特征,设计了一个密集膨胀图卷积模块,通过设置不同的膨胀率增大感受野,该模块中的密集连接形式,能够在有效利用特征的同时,加强特征间的信息传递。在所提出的网络结构中,基于多层感知器的离群点过滤模块,通过利用上下文标准化过滤掉不匹配的点对。在该网络中,匹配点云所需要的转换参数,利用奇异值分解模块获取。在三个广泛使用的数据集ModelNet40、ShapeNetCore与RealData上的实验结果,验证了所提出网络的有效性。Due to some unfavorable factors,such as irregular structure,noise and outliers in non-Euclidean space,how to accurately register partial point clouds is still a challenging task.For this task,an effective network is proposed for partial point cloud registration.In order to effectively extract the fine-grained features of local point cloud,a dilated graph convo-lution block with dense connection is devised by setting different expansion rates to increase the receptive field.The dense connection can effectively utilize the features and enhance the information transfer between features.Moreover,in the proposed network,an outlier filtering module based on multi-layer perceptron is explored to filter out mismatched point pairs by using context normalization.In this network,the transformation parameters for point cloud matching are derived via a singular value decomposition module.Experimental results on three widely used datasets,i.e.ModelNet40,ShapeNetCore and Real Data,demonstrate the effectiveness of the proposed network.

关 键 词：膨胀图卷积 密集连接 离群点过滤 点云配准 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]