Research On Feature Matching Algorithm Of Sequence Image

Image matching based on local features is one of the core technologies in the field of computer vision.It is also a key step in the following vision fields,such as image registration,3D reconstruction,image mosaic,moving object tracking and so on.Different from the general scene,the feature matching algorithm between sequence images needs to be more resistant to the change of perspective and illumination.Especially in some special scenes,such as large perspective scene,quasi dense scene and weak texture scene,the traditional feature matching algorithm has some defects,and the matching performance needs to be improved.Therefore,this paper proposes the corresponding feature matching algorithm for three kinds of special scenes in sequence image matching to further improve the matching performance.The main work of this paper is as follows:1.Aiming at the problems of low time efficiency and slow matching speed of existing feature matching algorithms in large view scenes,a feature point matching algorithm based on affine transformation space is proposed.Firstly,the affine change space is constructed to simulate the change of view angle to obtain affine invariance;secondly,the effective region is divided to avoid the feature point detection in the invalid region;in the feature description stage,the orb algorithm is integrated into the affine transformation space,and the gradient contrast information of multiple directions in the feature point sampling region is fused to obtain the final binary descriptor.Experiments on large view data sets and sequence images show that the algorithm has better matching effect in large view scenes,and has more advantages in time efficiency.2.In order to improve the accuracy of quasi dense matching and the number of matching points,a sparse matching algorithm based on Improved SIFT and a quasi dense matching algorithm based on multiple constraints are proposed.In the sparse matching stage,the Fast corner detection algorithm with adaptive threshold is used to detect the feature points.and the gradient-gray composite descriptor combines gradient contrast information and gray contrast information in SIFT descriptor;in the quasi dense matching stage,the sparse matching result is used as the seed point pair.and the quasi dense matching results are obtained according to certain synchronous growth strategy through the constraints of zero mean mutual normalization coefficient(ZNCC),confidence and disparity gradient.Experiments on Oxford dataset show that the sparse matching algorithm proposed in this paper has good resistance to various image changes such as view angle,illumination and scale;SFM reconstruction experiments on sequential images show that the quasi dense matching algorithm proposed can effectively improve the accuracy and consistency of sequential image matching.3.In order to solve the problem of low distinguishability and poor matching performance of descriptors in weak texture scene,a binary descriptor based on local entropy and multi-scale feature fusion is proposed.Firstly,a 4-dimensional dog space based on gray HSV information is established to add color information to the descriptor;secondly,orb features at multiple scales are obtained by pooling main dimensions;finally,the final binary descriptor is obtained based on local entropy of feature points and multi-scale fusion strategy.Experiments on Oxford dataset show that the our binary descriptor has good resistance to the changes of illumination and visual angle.Experiments on the sequence images with weak texture scenes show that our descriptor can effectively improve the matching accuracy in weak texture scenes,and has better matching effect and robustness.