Improvement And Evaluation Of Point Feature Extraction And Registration Algorithms Based On Visual SLAM

For visual SLAM(Simultaneous Localization And Mapping)technology,the image contains wealthy environmental information,and some characteristic information can be obtained through the processing of the image,such as points,lines,surfaces,and other characteristics so that image processing can obtain important external information.Point features,as the essential feature in an image,are also one of the most commonly used feature types.Image feature point detection and matching technology have a comprehensive application in computer vision applications.So far,it has been widely used in driverless,three-dimensional reconstruction and medical diagnosis,and other technical fields.After decades of development,image feature point detection technology has matured,and many classical algorithms have been derived.Later,many scholars preprocessed the image when extracting the feature points to make the feature point distribution more uniform.However,there are still many problems in the feature extraction and registration algorithm based on visual SLAM images,such as the small number of matches and low accuracy.This paper mainly studies the detection and matching technology of ORB(Oriented FAST and Rotated BRIEF,directional accelerated block test feature points and rotated binary stable,independent basic features)feature points,improves the algorithm,and evaluates the main work and results as follows:(1)Analyze and evaluate various feature extraction and matching point screening algorithms,and compare and analyze each feature extraction method through the computer platform to obtain various methods’ performance and advantages and disadvantages.(2)Aiming at the aggregation phenomenon of ORB feature point extraction,the feature point homogenization method used in the ORB-SLAM2 program was studied.In the process of feature point detection,the average distribution method and the quadtree method are fused to make the feature detection more uniform.In the feature description stage,the grayscale information of the feature point itself is used to fuse the grayscale information of the feature point itself in r BRIEF(rotated Binary Robust Independent Elementary Feature)to describe the feature point and increase the number of matches.In the stage of feature point uniformity detection,a stepped meshing method is proposed to improve the accuracy of feature point uniformity evaluation.The experimental results show that the homogenization evaluation results are obviously contrasted.Compared with the average distribution uniformity before and after the improvement,the uniformity was reduced by 6.55%.(3)The screening methods of feature matching point pairs are given,and several screening methods are compared and analyzed.The filtering algorithm of GMS(Grid-based Motion Statistics)is improved,which is evaluated using polar constraints.The root means square error after feature point mapping is calculated to evaluate the matching accuracy of feature points.Experiments show that the improved feature description method matches more feature points.Through the number of matching points filtered by the combination of Hamming distance and RANSAC(RANdom SAmple Consensus),the improved descriptor is,on average,33.8% higher than the number of matching points of the r BRIEF descriptor.Then the RMS error is used to evaluate the GMS matching results.The improved method shows lower error than the unimproved method.