杂乱室外场景下基于改进LoFTR的图像特征匹配方法  

Improved LoFTR Based Image Feature Matching Method in Cluttered Outdoor Scenes

作　　者：秦伦明 李博 崔昊杨 边后琴 王悉[2] QIN Lunming;LI Bo;CUI Haoyang;BIAN Houqin;WANG Xi(College of Electronics and Information Engineering,Shanghai University of Electric Power,Shanghai 201306,China;School of Electronic and Information Engineering,Beijing Jiaotong University,Beijing 100044,China)

机构地区：[1]上海电力大学电子与信息工程学院,上海201306 [2]北京交通大学电子信息工程学院,北京100044

出　　处：《电子信息对抗技术》2024年第6期52-59,共8页Electronic Information Warfare Technology

基　　金：国家自然科学基金面上项目(62073024)。

摘　　要：针对图像特征匹配过程中由于杂乱的室外场景或待匹配目标被物体遮挡而产生的外点导致匹配精度低及鲁棒性差等问题,提出了一种融合基于密度的带噪空间聚类算法(Density Based Spatial Clustering of Applications with Noise,DBSCAN)与Two-stage策略的改进LoFTR的图像特征匹配方法D2S-LoFTR。首先将原始图像1/8维度的特征匹配作为初始匹配结果,使用基于空间密度的DBSCAN算法对其特征进行聚类,提取最优匹配对的同时滤除由外点造成的误匹配。接着裁剪出由聚类得到的两幅原始图像中的共视区域,使用卷积层注意力模块(Convolutional Block Attention Module,CBAM)对其特征重构后进行二次匹配,将匹配结果与初始匹配进行融合以增强匹配的准确性。在室外数据集Megadepth上的实验结果表明,D2S-LoFTR的平均特征匹配率达到93.47%,与LoFTR相比提升1.91%,在旋转误差阈值为5°,10°,20°情况下的相对位姿估计累计曲线下面积(Area Under the cumulative Curve,AUC)分别为55.12%、71.03%、82.02%,分别提升2.32%,1.84%,0.84%,证实所提方法能够更好地适应杂乱室外场景下的图像特征匹配任务。Aiming at the problems of low matching accuracy and poor robustness due to the outliers generated by cluttered outdoor scene or the target to be matched being occluded by an object,an improved LoFTR image feature matching method D2S-LoFTR is proposed by integrating the density based spatial clustering of applications with noise(DBSCAN)and the two-stage strategy.Firstly,the feature matching of 1/8th dimension of the original image is used as the initial matching result,and the spatial density based DBSCAN algorithm is used to cluster the features to extract the optimal matching pairs while filtering out the mis-matches caused by outliers.Then the co-visual regions in the two original images obtained from clustering are cropped out,and secondary matching is performed after reconstructing their features using the convolutional block attention module(CBAM)attention module,and the matching results are fused with the initial matching to enhance the accuracy of the matching.Experimental results on the outdoor dataset Megadepth show that the average feature matching rate of D2S-LoFTR reaches 93.47%,which is an improvement of 1.91%compared with LoFTR,and the area under the cumulative curves(AUCs)of relative position estimation at rotation error thresholds of 5°,10°,and 20°are 55.12%,71.03%,and 82.02%,which are an improvement of 2.32%,1.84%,and 0.84%,respectively,confirming that the method can be better adapted to the task of matching image features in cluttered outdoor scenes.

关 键 词：外点 共视区域 DBSCAN CBAM LoFTR 图像特征匹配 

分 类 号：TN971[电子电信—信号与信息处理]