基于邻域一致性的点云场景流传播更新方法  

Point Cloud Scene Flow Propagation Update Method Based on NeighborhoodConsistency

作　　者：郑晗 王宁[1] 马新柱 张宏[1] 王智慧[1] 李豪杰[1] Zheng Han;Wang Ning;Ma Xinzhu;Zhang Hong;Wang Zhihui;Li Haojie(International School of Information Science and Engineering,Dalian University of Technology-Ritsumeikan University,Dalian,Liaoning 116000;Department of Electrical and Information Engineering,The University of Sydney,Sydney,Australia 2006)

机构地区：[1]大连理工大学-立命馆大学国际信息与软件学院,辽宁大连116000 [2]悉尼大学电气与信息工程系,澳大利亚悉尼2006

出　　处：《计算机研究与发展》2023年第2期426-434,共9页Journal of Computer Research and Development

基　　金：国家自然科学基金项目(61976038,61932020,61772108,U1908210);中央高校基本科研业务费专项资金(DUT20GF18)。

摘　　要：场景流是连续动态场景之间的3D运动场,广泛应用于机器人技术和自动驾驶任务.现有方法忽略了点云点的相关性,仅关注源点云和目标点云逐点的匹配关系,由于匹配关系完全依赖于点云数据的特征信息,导致在局部特征信息不足的点上准确估计场景流仍然存在挑战.根据源点云相邻点具有相关性的特性,提出NCPUM(neighborhood consistency propagation update method)方法,在邻域内将场景流从高置信度点向低置信度点传播,从而优化局部特征信息不足点的场景流.具体来说,NCPUM包含2个模块:置信度预测模块,根据场景流先验分布图,预测源点云逐点的置信度;场景流传播模块,根据局部区域一致性的约束更新低置信度点集的场景流.NCPUM在合成数据集Flyingthings3D和真实驾驶场景数据集KITTI上进行评估,准确度上达到了国际先进水平.由于邻域一致性更符合真实激光雷达场景的先验假设,因此在KITTI数据集上的提升更加明显.Scene flow is a 3D motion field between continuous dynamic scenes,which is widely applied in robotics and autonomous driving tasks.Existing methods ignore the correlation of point cloud points and focus only on the point-by-point matching relationship between the source point cloud and target point cloud,which is still challenging to estimate the scene flow accurately at the points with insufficient local feature information,because the matching relationship depends entirely on the feature information of the point cloud data.Considering the correlation property of the source point cloud’s local regions,the NCPUM(neighborhood consistency propagation update method)is proposed to propagate the scene flow from high-confidence points to low-confidence points in local regions,so as to optimize the scene flow at the points with insufficient local feature information.Specifically,NCPUM consists of two modules:the confidence prediction module,which predicts the confidence of the source point cloud according to the priori distribution map of scene flow;the scene flow propagation module,which updates the scene flow of the low confidence point set based on the local area consistency constraint.We evaluate NCPUM on both challenging synthetic data from Flyingthing3D and real Lidar scans from KITTI,and experiment results outperform by previous methods a large margin in accuracy,especially on KITTI dataset,because the neighborhood consistency is more applicable with the a priori assumptions of real Lidar scans.

关 键 词：场景流 点云 邻域一致性 置信度 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]