Image Feature Extraction Algorithm In Visual SLAM

Visual simultaneous localization and mapping(v SLAM)is a key technology for robots to autonomously locate and perceive the environment.Image feature extraction is an important part of the front-end in v SLAM,which directly affects the positioning accuracy and mapping accuracy of the robot.This thesis focuses on the problem of image feature extraction,and researches on manual features and features based on deep learning theory.It aims to propose a feature extraction algorithm with strong robustness and high matching accuracy,so as to provide correct matching feature point pairs for the subsequent links of v SLAM.The main work of this paper is as follows:To address the problem of low matching accuracy of the ORB algorithm,an improved ORB algorithm based on affine transformation is proposed.First,after detecting FAST points,the descriptors under different affine transformations are extracted from the feature points,and then the stable bits of the descriptor are extracted as feature descriptors for feature matching to eliminate the influence of unstable bits on the descriptor distance.To further improve the accuracy of feature matching,we use an improved F-SORT screening algorithm after feature matching.First,the matched features are grouped in order,and then the matched feature pairs are further filtered according to the angle,scale and distance information of the feature points.Experiments on VGG dataset and Zu Bud dataset show that the proposed algorithm has better performance than ORB algorithm in matching accuracy and number of correct matches.Aiming at the limitation that manual features cannot fully describe image information,a feature extraction algorithm based on feature fusion and adaptive weight under deep learning is proposed.First,the VGG network is used to extract the local descriptors of the feature points,and the shallow and deep layers are used for feature fusion,and then Net VLAD is used to further extract the local descriptors as the regional descriptors of the feature points.In the feature matching process,the shallow descriptors and the deep descriptors are weighted according to the similarity of the regional descriptors.In the process of model training,the triple loss function is used to improve the accuracy of feature point matching.Experiments on the HPatches dataset and RDNIM dataset show that the proposed algorithm has good matching performance and is suitable for challenging conditions such as large changes in views and day-night situations.The research in this paper is an important part of the v SLAM front-end for the National Natural Science Foundation project《Collaborative mapping and localization of heterogeneous robots in satellite signal denied environments》.The extracted and matched feature points provide data for the pose estimation of the robot in the v SLAM front-end.The achievements in this thesis can be used not only for v SLAM pose estimation,but also for face recognition,target tracking and other fields,which have theoretical and practical significance for promoting the v SLAM technology in robotics applications.