摘要
已有的实时定位与地图构建(simultaneous localization and mapping, SLAM)方案采用的特征点匹配方法普遍会受视角变化的影响使得特征点的匹配比较困难,进而干扰到特征点匹配的精度,最终影响到三维(three-dimensional, 3D)点云地图构建以及相机运动位姿估计的精度。为此,本文提出一种基于注意力机制的特征点匹配网络的SLAM方法。相比于现有的SLAM方法,本文将SLAM中视觉里程计模块的特征点匹配的方法替换成了一个全新的、基于注意力机制的特征点匹配网络的特征点匹配方法,并和传统的特征点提取方法做了一个全新的特征点提取与匹配的组合,形成了一个新的视觉里程计,进而形成了一个新的SLAM方法。首先,通过传统的特征点提取算法进行特征点的提取,对提取的特征点及描述子向量进行编码,通过图注意力神经网络进行学习得到匹配描述子,根据匹配描述子创建得分矩阵,采用最优传输算法求解最优得分矩阵,计算得到最优匹配点对,到这里就完成了特征点提取与匹配的整个过程;基于匹配点对完成相机的定位、建图和回环检测。本文采用KITTI公开数据集进行实验,实验结果表明采用基于注意力机制特征点匹配网络的SLAM方案,在视角变化不稳定的情况下,相机运动轨迹误差和相机位姿估计误差的精度明显有所提升。
The current feature point matching method for simultaneous localization and mapping(SLAM) is generally affected by the change of perspective, which makes the matching of feature points difficult, which in turn deteriorates the accuracy of feature point matching, and ultimately influences the construction of three-dimensional(3 D) point cloud maps and the estimation accuracy of camera motion pose.For this reason, this paper presents an attention based on feature point matching network for SLAM.The innovation of this article is that compared with the existing SLAM and we replaces the feature point matching method of the visual odometer module in SLAM with an attention based on feature point matching network for feature point matching.And we make a new combination of feature point extraction and matching with the traditional feature point extraction method to form a new visual odometer and a new SLAM.Firstly, we encode the extracted feature points and descriptor vectors and we learn through the graph attention neural network to obtain matching descriptors.Then we create a score matrix based on the matching descriptors and use the optimal transmission algorithm to solve the optimal score matrix.In the end we calculate the optimal matching point pair and complete camera positioning, mapping, and loop detection based on the optimal matching point pairs.The experimental results show that when the viewing angle is unstable, an attention based on feature point matching network for SLAM can significantly improve the accuracy of the camera′s trajectory and the estimation accuracy of camera motion pose.
作者
俎晨洋
刘凤连
汪日伟
ZU Chenyang;LIU Fenglian;WANG Riwei(Key laboratory on Computer Vision and Systems,Ministry of Education of China,Key laboratory on Intelligence Computing and Novel Software Technology of the City of Tianjin,Tianjin University of Technology,Tianjin 300384,China;WenZhou University of lechoology,WenZhou,Zhejiang 325035,China)
出处
《光电子．激光》
CAS
CSCD
北大核心
2022年第1期14-22,共9页
Journal of Optoelectronics·Laser
基金
国家自然科学基金(62020106004)资助项目。
