Xfeat model-based visual SLAM localization algorithm in low texture scenes

Visual SLAM technology is an important research direction in the fields of mobile robots, low-altitude economy, automatic driving, and unmanned aerial vehicles. However, ORB algorithm has limited feature point extraction ability in low texture scenes, to address this problem, based on ORB-SLAM3 algorithm, XFeat deep learning model with context-awareness is used to replace ORB algorithm, to enhance the feature point extraction ability in low texture regions, and XFeat-SLAM algorithm is proposed to replace ORB-SLAM3. Comparison experiments of positioning accuracy are conducted on the TUM public dataset, and the results show that both algorithms, ORB-SLAM3 and XFeat-SLAM, can extract effective feature points in texture-rich scenarios, and in most sequences, the XFeat-SLAM algorithm is superior in positioning accuracy, with a maximum improvement of 5.16% in absolute trajectory error (ATE), a minimum improvement of 1.02%, and the relative position error (RPE) is improved by a maximum of 6.38% and a minimum of 2.06%; in low texture environments, XFeat-SLAM can extract a sufficient number of feature points to be able to form a complete running trajectory, with an ATE of 0.02m~0.12m and an RPE of 0.01m~0.04m, whereas ORB-SLAM3, due to its reliance on the traditional ORB-SLAM3, which relies on the traditional ORB algorithm, often fails to complete the tracking in feature-poor regions due to insufficient number of extracted feature points.