Research On Image Matching And Loop Detection Method Based On Point And Line Feature Fusion In Visual SLAM

In recent years,the rapid development of computer vision technology and sensor technology has made SLAM technology widely used in AR,robots,unmanned aerial vehicles,unmanned driving and other fields.However,in the application of feature point-based visual SLAM technology in indoor structured artificial environment,there are often problems of inaccurate system pose estimation and positioning tracking failure due to insufficient collection of feature points and uneven distribution.At the same time,in such environments There are often very rich line textures.These texture information can be described by line features.Based on this,this paper adds line feature information in the environment to the visual SLAM system,and proposes a method for image matching and loop detection based on the fusion of point and line features in visual SLAM.This method can effectively improve the positioning accuracy and robustness of the visual SLAM algorithm.The main research contents of this paper are as follows:(1)An image matching method that integrates point and line features.First extract the ORB point feature in the image and calculate its corresponding descriptor.At the same time,the image is preprocessed through the gradient density filter to remove the area with too high local line feature density,and the rest area uses the LSD algorithm to extract the line feature and calculate its Corresponding descriptors,and then group the line segments that conform to the angle,horizontal coordinate,and vertical coordinate closeness.By constructing a plane rectangular coordinate system,comprehensively considering the angle of the line segment to be merged and the coordinate information between each end point,the merge of the same group of line segments is completed.Then,the Hamming distance between each feature descriptor is used to measure the similarity between features.Finally,the brute force matching method is used for feature matching,and the wrong matching pairs are eliminated by the RANSAC algorithm.(2)Loop detection method that integrates point and line features.In this paper,the method of bag-of-words model based on feature points is extended,line features are also taken into consideration,and a new way of constructing vocabulary tree is proposed.First extract the ORB point feature and LSD line feature of the image,and calculate the corresponding descriptor.In order to ensure the uniformity of the word distribution in the image,the Kmeans++algorithm is used to cluster the features.In order to distinguish the point feature and the line feature,the dictionary is constructed In the tree process,the flag bit of the brief descriptor is set to 0,and the flag bit of the LBD descriptor is set to 1.Then,combining the actual characteristics of different scenes,set the weights of point features and line features reasonably.In order to distinguish the importance of words,this article uses the TF-IDF model to calculate the weight of each word.The similarity between the images is expressed by calculating the norm of the feature vocabulary pack vector between the images.When the similarity is high enough,it can be judged that the two frames of images may be a loop.Provide more pose constraints for map construction,reduce accumulated errors,and improve the composition accuracy of the system.(3)Experiment with public data sets to verify the feasibility and effectiveness of the algorithm in this paper.Select six classic scenarios in the TUM data set,compare the algorithm in this paper with the ORB-SLAM algorithm,draw the ATE trajectory distribution diagram in each scenario,and compare and analyze the maximum error,minimum error and root mean square error of the experimental ATE and RPE.The experimental results show that the algorithm in this paper has smaller root mean square error and maximum error compared with ORB-SLAM algorithm.Compared with ORB-SLAM algorithm,the algorithm in this paper has better positioning accuracy and robustness.The algorithm in this paper is similar to ORB-SLAM.Compared with the ORB-SLAM algorithm,it takes less time in the loop detection link,and basically the same time in the feature extraction,feature matching and pose estimation links,which meets the real-time requirements of the SLAM system.In summary,after adding the constraint of line features in the image matching and loop detection links of the visual SLAM system,the algorithm in this paper has improved the accuracy and robustness of the system's pose and has improved the visual SLAM algorithm in the indoor structured artificial environment.The following applications have a positive effect.