同步定位与建图特征提取和匹配算法研究

Research on Simultaneous localization and mapping feature extraction and matching algorithm

针对传统特征提取算法提取的特征点集中、重复率低的问题,提出了一种基于深度学习的特征提取网络模型,首先通过编码网络对图像进行特征压缩,分别使用特征点提取网络和描述子网络,得到图像的特征点和描述子,获得数量足够、分布均匀、具有代表性的特征点。之后采用图神经网络和注意力机制匹配特征点,将特征点和描述子编码成一个特征匹配向量,传入最佳匹配层,得到较好的匹配关系,从而估算出运动位姿的变化,提高了同步定位与建图(SLAM)系统的定位精度。在公开数据集上,分别用尺度不变特征变换算法(SIFT)、快速特征点提取和描述算法(ORB)和深度学习的方法作为角点检测器,对视觉里程计(VO)进行了分析和比较。实验结果表明,改进后的SLAM系统定位精度提高了33.56%。

Aiming at the problems of feature point concentration and low repetition rate extracted by traditional feature extraction algorithms,a feature extraction network model based on deep learning is proposed,which compresses the image features through the encoding network,and the feature points and descriptors of the image are obtained by using the feature point extraction network and descriptors network respectively,so as to obtain sufficient,evenly distributed and representative feature points.Then,the graph neural network and attention mechanism are used to match the feature points,the feature points and descriptors are encoded into a feature matching vector,which is introduced into the best matching layer to obtain a better matching relationship,so as to estimate the changes of motion pose and improve the positioning accuracy of Simultaneous Localization and Mapping(SLAM).The Visual Odometry(VO)is analyzed and compared with Scale-Invariant Feature Transform(SIFT),Oriented FAST and Rotated BRIEF(ORB)and methods of deep learning as corner detectors on public dataset.The experimental results show that the positioning accuracy of the improved SLAM system is improved by 33.56%.