Research On Key Technologies For Feature Points In Multi-View 3D Recovery

Feature points are the basis of multi-view 3D recovery,and can be extracted and matched to achieve image correspondence and geometric information from different views.Therefore,by extracting key points or points of interest with significant features in two images for matching,a correspondence can be established between the two images,and the accuracy and speed of their extraction and matching directly affect the accuracy and real-time performance of subsequent tasks.However,the number of extracted feature points is low and a large number of false matches can be generated due to specific environments such as low texture and illumination.When using RANSAC(Random Sample Consensus),all samples need to be randomly selected for iterative solving,which greatly affects the matching speed,and the presence of a large number of false matches in the initial matching can be detrimental to the matching accuracy of multi-level matching.To this end,this thesis has further investigated the feature extraction and matching methods,and the main results achieved can be summarized as follows:Firstly,an improved ORB(Oriented FAST and Rotated BRIEF)algorithm is proposed to extract feature points from capsule endoscopy images for the problem that it is difficult to extract feature points in regions with insignificant changes in capsule endoscopy images using a fixed threshold,and that redundancy of feature points occurs in significant regions.The method first uses variance coefficients to calculate adaptive thresholds in the feature point extraction stage to improve the algorithm’s ability to extract feature points in homogeneous regions.Then,the feature points are filtered using a quadtree approach to eliminate overconcentrated and overlapping feature points.Finally,BEBLID(Boosted Efficient Binary Local Image Descriptor)binary descriptors are used to construct binary descriptors with strong descriptions,and RANSAC is used to avoid mismatches.Secondly,to address the problem of the need to randomly select all the samples to iteratively solve the optimal model of RANSAC,which leads to a time-consuming method of rejecting false matches,a fusion of GMS(Grid-based Motion Statistics)and RANSAC is proposed to facilitate the speedup of visual SLAM(Simultaneous Localization and Mapping).The method first obtains the number of matching pairs in the neighborhood of a matching pair based on the GMS algorithm,and uses this as a confidence level to rank and group the matching pairs,and then uses the RANSAC algorithm to further reject false matches,converting the timeconsuming problem of randomly selecting all matching pairs into a problem of prioritizing the selection of samples from pairs with high confidence levels,and then implementing the iterative solution of the optimal model.Thirdly,to address the problem that when resampling-based methods are used in multi-level matching methods to reject false matches,the resampling-based methods are affected by the low initial matching accuracy,resulting in reduced matching accuracy.A new feature matching method based on a multi-stage fine matching strategy is established,known as the KTGP-ORB method.The method uses the similarity of local appearance of feature descriptors in Hamming space to generate the initial correspondence,and combines the local image motion smoothness constraint with the GMS algorithm to improve the initial matching accuracy;finally,the PROSAC(Progressive Sampling Consensus)algorithm is used to optimize the matching and obtain an accurate matching based on the global grey-scale information in Euclidean space.This thesis investigates the different problems of feature extraction and matching in different scenes,improving the problems of difficult feature extraction,slow matching and low matching accuracy in complex scenes,supporting the application of SLAM and SFM(Structure from Motion)techniques in a wider range of fields,and enriching the methodology and technical system of multi-view geometry.