CN103927785B - A kind of characteristic point matching method towards up short stereoscopic image data - Google Patents

Abstract

The present invention provides a feature point matching method for close-range photographic stereoscopic image data, which comprises the following steps: performing forward matching of feature points on image pairs with the same name using sub-regions and triangles as constraints to obtain a forward image point group with the same name, and The image of the same name performs reverse matching on the feature points with sub-regions and triangles as constraints, and obtains the reverse image point group of the same name, retains the matching results of the same-named image points in the forward image point group and the reverse image point group of the same name, and obtains the final match The result of the image point with the same name; the present invention adopts sub-regions and triangles as constraint conditions to carry out feature point detection, which greatly improves the time efficiency of feature point detection and improves the number of correct matching points, and has important applications in close-range photography stereoscopic image matching value.

Description

A feature point matching method for close-range photography stereo image data

technical field

The invention belongs to the field of digital close-range photography and relates to a feature point matching method.

Background technique

With the development of modern surveying and mapping technology, digital photogrammetry can provide data and support for digital earth, three-dimensional modeling of cities, etc. For the original data obtained by digital photogrammetry, the features of digital images are more important, including point features, line features and surface features. The method of extracting these features is the basis of image analysis and image matching, and various operators can be used. For example, Moravec operator, Forstner operator and Harris operator can all be used to extract feature points. For obvious objects in digital images, it is not only necessary to identify them, but also to determine their positions. The measurement of stereo pairs in photogrammetry is the basis for extracting three-dimensional information of objects. In digital photogrammetry, image matching is used to replace traditional manual observation to achieve the purpose of automatically determining the same-named image point. Image matching is essentially to identify image points with the same name between two or more images, which is the core problem of digital photogrammetry and computer vision. The matching method based on gray level is a relatively mature matching method, mainly including methods based on gray level similarity detection and least squares image matching, which are all based on the similarity of images with the same name.

Scale-invariant feature transform (Scale-invariant feature transform, SIFT) is a local feature descriptor proposed by David Lowe, which combines the steps of spot detection, feature vector generation, and feature matching search to achieve near real-time operating speed. The SIFT feature matching algorithm can deal with the matching problem in the case of translation, rotation, and affine transformation between two images, and has a strong matching ability. However, when acquiring image data, due to the influence of factors such as precipitation reflection, soil movement, high-speed camera focal length setting and calibration accuracy, shooting angle, and scene layout, the image data is prone to blurring, inconspicuous features, and blurred textures. At this time, if the conventional image matching method is used, there will be fewer correct matching points and more wrong matching points.

Contents of the invention

The purpose of the present invention is to provide a feature point matching method for close-range photography stereo image data that can increase the number of correct matching points, reduce the number of wrong matching points, and improve the time efficiency of feature point detection.

To achieve the above object, the solution of the present invention is:

A method for matching feature points of close-range photographic stereoscopic image data, comprising the following steps:

(1) Use parallel photography to shoot the scene to be matched, select two images taken by two cameras at the same time as the first image and the second image, and select the first image to be matched in the first image Region, in the second image and select the second region to be matched corresponding to the first region to be matched;

(2), utilize the SIFT algorithm to detect respectively the feature point of the first area to be matched and the feature point of the second area to be matched, and take the feature points in the first area to be matched as a benchmark to separate the first area to be matched and the second area to be matched The matching area is divided into multiple one-to-one sub-areas;

(3), each sub-area of the first area to be matched is used as a constraint condition, and the nearest neighbor distance algorithm is used to carry out the forward matching of the feature points from the sub-area of the first area to be matched to the corresponding sub-area of the second area to be matched , and finally match the image points with the same name in each sub-region;

(4), with the image point of the same name of each sub-area of the first region to be matched and the feature point positioned on the edge of the sub-area as three vertices of the triangle to divide the sub-area and build a triangular network;

(5), with each triangle in the triangle network of step (4) gained as constraint condition, utilize the nearest neighbor distance algorithm to carry out by the triangle of the first area to be matched to the feature point circular forward matching of the corresponding triangle of the second area to be matched , to get the positive image point group of the same name;

(6), each sub-area of the second area to be matched is used as a constraint condition, and the nearest neighbor distance algorithm is used to carry out reverse matching of feature points from the sub-area of the second area to be matched to the corresponding sub-area of the first area to be matched, Finally, the image points with the same name in each sub-region are matched;

(7), with the image point of the same name of each sub-area of the second region to be matched and the feature point positioned on the edge of the sub-area as three vertices of the triangle to divide the sub-area and build a triangular network;

(8), with each triangle in the triangle network of step (7) gained as constraint condition, utilize the nearest neighbor distance algorithm to carry out by the triangle of the second area to be matched to the feature point circular reverse matching of the corresponding triangle of the first area to be matched, Obtain the reverse image point group of the same name;

(9). According to the forward image point group with the same name and the reverse image point group with the same name, a correctly matched image point result with the same name is obtained.

In step (2), according to the respective slopes and slopes of the first area to be matched and the second area to be matched, the feature points that are visually identical in the first area to be matched and the second area to be matched are used as the segmentation points of the sub-areas , completely divide the first to-be-matched region and the second to-be-matched region into several corresponding sub-regions. Further, adjacent sub-regions of the first to-be-matched region or the second to-be-matched region partially overlap each other.

The feature point circular forward matching in the step (5) specifically includes:

The first step is to set the preset number of cycles and the triangle size threshold;

Second step, each triangle in step (4) gained triangular network is used as the first-order triangle, takes the first-order triangle as constraint condition, utilizes nearest neighbor distance algorithm to carry out by the first-order triangle of the first area to be matched to the first-order triangle 2. The feature points of the corresponding triangles in the area to be matched are forward matched to obtain the same-named image points of each first-level triangle, n is set to 1, and m is set to 0;

The 3rd step, n=n+1, utilize the homonym image point of correct matching in the n-1th grade triangle and the division point of the subregion as three vertices of the triangle to construct the nth grade triangle;

The fourth step is to use the nth level triangle as a constraint condition, and use the nearest neighbor distance algorithm to carry out forward matching from the nth level triangle of the first to-be-matched area to the feature point of the corresponding triangle in the second to-be-matched area, and through multiple The cyclic RANSAC algorithm eliminates the incorrectly matched feature points, and obtains the correctly matched image points with the same name, m=m+1;

The fifth step is to determine whether m is not less than the preset number of cycles and whether the size of the nth-level triangles is smaller than the triangle size threshold. When m is not less than the preset number of cycles and the size of the n-th-level triangles is smaller than the triangle size threshold, get Forward the point group of the same name, otherwise return to the third step,

Among them, n is the level of the triangle, and m is the number of cycles.

The feature point circular reverse matching in the step (8) specifically includes:

The first step is to set the preset number of cycles and the triangle size threshold;

Second step, each triangle in step (7) gained triangular network is used as the first-order triangle, takes the first-order triangle as constraint condition, utilizes nearest neighbor distance algorithm to carry out by the first-order triangle of the second area to be matched to the first-order triangle The feature points of the corresponding triangles in the area to be matched are reversely matched to obtain the same-named image points of each first-level triangle, n' is set to 1, and m' is set to 0;

The 3rd step, n'=n'+1', utilize the homonym image point of correct match in the n'-1th grade triangle and the dividing point of the subregion as the three vertices of the triangle to construct the n'th grade triangle;

The fourth step is to use the n'th level triangle as a constraint condition, and use the nearest neighbor distance algorithm to carry out reverse matching of the feature points from the n'th level triangle in the second to-be-matched area to the corresponding triangle in the first to-be-matched area, and through multiple The sub-cycle RANSAC algorithm eliminates the mismatched feature points, and obtains the correctly matched image points with the same name, m'=m'+1;

The fifth step is to judge whether m' is not less than the preset number of cycles and whether the size of the n'th level triangle is smaller than the triangle size threshold, when m' is not less than the preset number of cycles and the size of the n'th level triangle is smaller than the triangle size When the threshold is reached, get the reverse image point group with the same name, otherwise return to the third step,

Among them, n' is the level of the triangle, and m' is the number of cycles.

Owing to adopting said scheme, the beneficial effect of the present invention is:

First of all, this method sequentially uses sub-regions and Delaunay triangles as constraints to detect feature points, which greatly improves the time efficiency of feature point detection; secondly, this method repeatedly uses the cyclic RANSAC algorithm to eliminate incorrectly matched feature points, thereby reducing the number of culling points. The probability of the correct image point with the same name; in addition, this method uses the SIFT operator to perform matching based on features, which does not involve the internal and external orientation elements of the image, and avoids the impact of unknown camera parameters or large errors in camera calibration parameters on the accuracy of the matching results; finally, This method uses level-by-level constraint feature point regions for matching, which avoids mismatching caused by similar features.

Description of drawings

FIG. 1 is a flow chart of a feature point matching method for close-range photography stereo image data in an embodiment of the present invention.

FIG. 2 is a flow chart of forward matching of feature points performed in a loop in an embodiment of the present invention.

Fig. 3 is a flow chart of the reverse matching of feature points performed in a loop in the embodiment of the present invention.

detailed description

The present invention will be further described below in conjunction with the embodiments shown in the accompanying drawings.

Example

This embodiment provides a feature point matching method for close-range photographic stereoscopic image data, which mainly uses sub-regions and Delaunay triangles as constraint conditions, uses SIFT operator to detect and match feature points, and is accompanied by multiple rounds of RANSAC calculations. Random Sample Consensus (Random Sample Consensus) deletes wrongly matched pixels with the same name, thereby increasing the number of correct matching points and reducing the number of wrong matching points.

As shown in Figure 1, the feature point matching method for close-range photography stereoscopic image data in the present embodiment includes the following steps:

The first step is to obtain the image pair with the same name and select the area to be matched, including:

Use a high-speed camera to shoot the same scene to be matched in parallel photography to obtain multiple digital images, and select two images taken by two high-speed cameras at the same time as the first image and the second image respectively (that is, form the image pair with the same name). In order to improve the time efficiency of feature point detection, select the area to be matched from the first image as the first area to be matched according to specific needs, and select the area corresponding to the first area to be matched from the second image as the second area to be matched area.

In the second step, the feature points of the first region to be matched and the feature points of the second region to be matched are respectively detected by using the SIFT algorithm. The SIFT algorithm includes feature point detection and image point matching of the same name. Feature point detection is to search out the local extreme points of the image in the scale space of the image, and then remove the low-contrast extreme points and unstable edge response points, so that Determine the feature points of the image, then divide the image into blocks around the feature points, calculate the gradient histogram in each block, and generate a unique vector descriptor, that is, a 128-dimensional feature descriptor, to describe the position of each feature point , scale and orientation information.

In the third step, based on the feature points in the first region to be matched, the first region to be matched and the second region to be matched are divided into a plurality of subregions, so that the subregions of the first region to be matched and the second region to be matched There is a one-to-one correspondence between subregions of regions. When segmenting, it is necessary to comprehensively consider the characteristics of the slope and aspect of the first to-be-matched area and the second to-be-matched area, and use the feature points that are visually identical in the first to-be-matched area and the second to-be-matched area as sub-groups. The segmentation point of the area completely divides the first to-be-matched area and the second to-be-matched area into several corresponding sub-areas. During segmentation, it is necessary to ensure that the adjacent sub-regions of the first to-be-matched region or the second to-be-matched region partially overlap each other, so as to ensure successful matching of feature points at the junction of the sub-regions.

The fourth step is to use each sub-area of the first area to be matched as a constraint condition, and use the nearest neighbor distance algorithm to perform one-way matching of feature points from the sub-area of the first area to be matched to the corresponding sub-area of the second area to be matched (forward matching), and remove the incorrectly matched feature points through the multiple-cycle RANSAC algorithm, thereby matching the image points with the same name in each sub-region of the first to-be-matched region and the second to-be-matched region.

In this step, the matching of the same-named image points in the SIFT algorithm adopts the nearest neighbor distance algorithm. The nearest-neighbor distance algorithm uses sub-regions as constraints, which can limit the search range of the same-named image points, which not only improves the time efficiency of the same-named image point matching, It also significantly reduces the false match rate. Under this constraint, according to the detected feature points in each sub-area, the feature points of each sub-area are matched using the nearest neighbor distance algorithm, that is, the Euclidean method of the feature point nearest to the sample feature point is used. The ratio of the distance to the Euclidean distance of the second nearest neighbor feature point is compared with the set threshold, and if the ratio is smaller than the threshold, it is considered as a feature point. However, for the matched feature points in each sub-region, there will be mis-matched results, and the mis-matched feature points can be eliminated through the multiple-cycle RANSAC algorithm.

The RANSAC algorithm is an algorithm that calculates the mathematical model parameters of the data based on a set of sample data sets containing abnormal data, and obtains effective sample data. The basic assumption of the algorithm is that the sample contains both correct data and abnormal data. When a set of correct data is given, there is a method that can calculate the model parameters that conform to these data. The RANSAC algorithm first randomly selects a subset in the sample data set and initializes it as a model, judges whether other data in the data set satisfy the model, and if so, forms a consistent set and calculates a new model until the end of the algorithm. According to the principle of the algorithm, it can be concluded that due to the different subsets of randomly selected data, the initial model constructed will be different, and some correct data are eliminated because they do not satisfy the model and are mistaken for wrong data. The sample data contained in the consistent set and the results obtained by the algorithm will also be different.

If only the SIFT operator is used to detect the feature points and match the same-named image points on the entire area to be matched, there will be many mismatching points in the matching result, which requires the use of the RANSAC operator to eliminate the mismatching points. If you run the RANSAC operator only once on the result, although there are only correct pixels with the same name in the result, a large number of correct matching points are also considered as mismatching points and are eliminated, so you need to cycle the RANSAC algorithm; when the cycle RANSAC algorithm reaches hundreds times, since the initial model of each cycle may be different, it can reduce the probability of eliminating the correct pixel with the same name.

The 5th step, with the same name image point that has been matched in each sub-area and the division point of sub-area as three apexes of triangle, respectively in each sub-area of the first area to be matched and each sub-area of the second area to be matched Construct the Delaunay triangle, and use the Delaunay triangle as the first-level triangle, and multiple first-level triangles together form the first-level triangular network.

The sixth step, with each first-level triangle in the first-level triangular network obtained in the fifth step as the initial constraint condition, use the nearest neighbor distance algorithm to carry out the corresponding triangle from the triangle in the first area to be matched to the corresponding triangle in the second area to be matched The feature point circular one-way matching (forward matching) and the construction of the nth level triangle finally get the forward homonymous image point group.

Wherein, as shown in Figure 2, the sixth step specifically includes the following sub-steps:

Step 1-1, set the preset number of cycles and triangle size threshold;

Step 1-2, with the first-level triangle obtained in the fifth step as the constraint condition, use the nearest neighbor distance algorithm to perform one-way matching from the first-level triangle in the first to-be-matched area to the feature point of the corresponding triangle in the second to-be-matched area (forward matching), and through multiple cycles of RANSAC algorithm to eliminate the mismatched feature points, to obtain the correct matching image points of the same name of each first-level triangle, the level n of the triangle is set to 1, and the value of the number of cycles m is set is 0;

Step 1-3, making the level of the triangle n=n+1, using the correctly matched image point of the same name in the n-1th level triangle and the division point of the sub-region as the three vertices of the triangle to construct the nth level triangle;

Step 1-4, with the nth level triangle as the constraint condition, the nearest neighbor distance algorithm is used to carry out the one-way matching (forward matching) of the feature points of the corresponding triangle of the second area to be matched from the nth level triangle of the first area to be matched ), and get rid of the mismatched feature points by multiple cycles of RANSAC algorithm, to obtain the correctly matched image point of the same name, the number of cycles m=m+1;

Step 1-5, judge whether the number of cycles m is not less than the preset number of cycles m ₀ and whether the size Tn of the nth-level triangles is smaller than the triangle size threshold T ₀ , when the number of cycles m is not less than the preset number of cycles and the n-th level triangle When the sizes of are smaller than the triangle size threshold, get the positive image point group with the same name, and then end the sixth step, otherwise return to step 1-3.

The seventh step is to use each sub-area of the second area to be matched as a constraint condition, and use the nearest neighbor distance algorithm to perform one-way matching of feature points from the sub-area of the second area to be matched to the corresponding sub-area of the first area to be matched (reverse matching), and remove the incorrectly matched feature points through multiple cycles of the RANSAC algorithm, thereby matching the image points with the same name in each sub-region of the first to-be-matched region and the second to-be-matched region.

The 8th step, with the same name picture point that has been matched in each sub-area and the dividing point of sub-area as three apexes of triangle, respectively in each sub-area of the first area to be matched and each sub-area of the second area to be matched A Delaunay triangle is constructed, and the Delaunay triangle is used as a first-level triangle, and a plurality of first-level triangles together form a first-level triangular network.

The ninth step, with each first-level triangle in the first-level triangular network obtained in the eighth step as the initial constraint condition, use the nearest neighbor distance algorithm to carry out the corresponding triangle from the triangle in the second area to be matched to the corresponding triangle in the first area to be matched The feature point cyclic one-way matching (reverse matching) and the construction of the n'th level triangle finally get the reverse image point group with the same name.

Wherein, as shown in Figure 3, the ninth step specifically includes the following sub-steps:

Step 2-1, setting the preset number of cycles and the triangle size threshold;

Step 2-2, with the first-level triangle obtained in the eighth step as a constraint condition, use the nearest neighbor distance algorithm to perform one-way matching from the first-level triangle in the second to-be-matched area to the feature point of the corresponding triangle in the first to-be-matched area (reverse matching), and eliminate the mis-matched feature points through the multiple-cycle RANSAC algorithm to obtain the correct matching image points of the same name of each first-level triangle, the level n' of the triangle is set to 1, and the value of the number of cycles m' is set is 0;

Step 2-3, making the level n'=n'+1 of the triangle, using the correctly matched image points of the same name in the n-1th level triangle and the division points of the sub-regions as the three vertices of the triangle to construct the n'th level triangle;

Step 2-4, with the n 'th level triangle as the constraint condition, utilize the nearest neighbor distance algorithm to carry out the one-way matching (reverse matching) of the feature points of the corresponding triangle of the first area to be matched from the nth level triangle of the second area to be matched ), and remove the mismatched feature points by multiple cycles of the RANSAC algorithm to obtain correctly matched image points of the same name, so that the number of cycles m'=m'+1;

Step 2-5. Determine whether the number of cycles m' is not less than the preset number of cycles m' ₀ and whether the size T' _n of the n'th level triangle is smaller than the triangle size threshold T' ₀ , when the number of cycles m' is not less than the preset When the number of cycles is repeated and the size of the n'th level triangles is smaller than the triangle size threshold, the reverse image point group with the same name is obtained, and then the ninth step is ended, otherwise, return to step 2-3.

Because image pairs with the same name can detect very dense feature points through the SIFT operator, visual observation reveals that there will be a large number of image points with the same name, after matching the same-named image points through the sub-region and triangular network constraints, and eliminating the mismatched points through the RANSAC operator , there are still many pixels with the same name that have not been successfully matched. Therefore, reverse matching can be used, that is, the reverse matching of feature points can be performed based on another image in the image pair with the same name.

In the tenth step, retain the matching result of the same-named pixel in the forward homonymous pixel group and the reverse homonymous pixel group, that is, obtain the final matching homonymous pixel result.

According to the method mentioned in this embodiment, use the stereo image data captured by the high-speed camera under the landslide simulation platform to verify its feasibility. When the matching method based on gray similarity is used instead of the method of this embodiment, only 422 pairs of image points with the same name are correctly matched for the entire landslide body, and only 12 image points with the same name are matched for the lower part of the landslide body. Using the method of this embodiment, with sub-regions and circularly constructed triangular networks as constraints, image pairs with the same name are used as reference images successively, matching is performed according to the SIFT algorithm, and incorrectly matched points are eliminated by the RANSAC algorithm for multiple cycles. After testing the data, the results show that 4,552 image points with the same name can be matched evenly in the image pair with the same name, and the matching number is 10 times that of the previous one, which greatly increases the number of correct matching points.

To sum up, under the influence of imaging model, external environment and other factors, the quality of image data acquired by high-speed cameras is poor. If SIFT and RANSAC algorithms are directly used, there will be few and uneven distribution of successfully matched pixels with the same name. ; And the present embodiment adopts subregion and triangular network as the constraint condition multiple cycles RANSAC algorithm and carries out the two-way matching of image pair of the same name, obtains image point data of the same name is many and evenly distributed, and reduces the wrong match point, in close-range photography three-dimensional It has important application value in image matching and three-dimensional model reconstruction.

The above description of the embodiments is for those of ordinary skill in the art to understand and use the present invention. It is obvious that those skilled in the art can easily make various modifications to these embodiments, and apply the general principles described here to other embodiments without creative efforts. Therefore, the present invention is not limited to the above-mentioned embodiments. Improvements and modifications made by those skilled in the art according to the disclosure of the present invention without departing from the scope of the present invention should fall within the protection scope of the present invention.

Claims (5)

1. the characteristic point matching method towards up short stereoscopic image data, it is characterised in that: comprise the steps:

(1), use the mode of parallel photography treat coupling scene shoot, select two cameras of synchronization shoot respectively two Width image, as piece image and the second width image, chooses the first region to be matched in described piece image, described In two width images and choose the second to be matched region corresponding with described first region to be matched；

(2) SIFT algorithm, is utilized to detect characteristic point and the spy in described second region to be matched in described first region to be matched respectively Levy a little, respectively described first region to be matched and described second are treated on the basis of the characteristic point in described first region to be matched Joining region segmentation is multiple subregions one to one；

(3), respectively with each sub regions in described first region to be matched about restraint condition, utilize nearest neighbor distance algorithm carry out by The subregion in described first region to be matched mates to the characteristic point forward of the corresponding subregion in described second region to be matched, finally Match the corresponding image points of each sub regions；

(4), with the corresponding image points of each sub regions in described first region to be matched and the characteristic point that is positioned on described subregion edge Three summits as triangle are split subregions and build triangular network；

(5), with each triangle in step (4) gained triangular network as constraints, nearest neighbor distance algorithm is utilized to carry out Forward coupling is circulated to the characteristic point of the corresponding triangle in described second region to be matched by the triangle in described first region to be matched, Obtain forward corresponding image points group；

(6), respectively with each sub regions in described second region to be matched about restraint condition, utilize nearest neighbor distance algorithm carry out by The subregion in described second region to be matched inversely mates, finally to the characteristic point of the corresponding subregion in described first region to be matched Match the corresponding image points of each sub regions；

(7), with the corresponding image points of each sub regions in described second region to be matched and the characteristic point that is positioned on described subregion edge Three summits as triangle are split subregions and build triangular network；

(8), with each triangle in step (7) gained triangular network as constraints, nearest neighbor distance algorithm is utilized to carry out Reverse coupling is circulated to the characteristic point of the corresponding triangle in described first region to be matched by the triangle in described second region to be matched, Obtain reverse corresponding image points group；

(9), the corresponding image points result correctly mated according to described forward corresponding image points group and described reverse corresponding image points group.

Characteristic point matching method towards up short stereoscopic image data the most according to claim 1, it is characterised in that: In step (2), according to described first region to be matched and the described second respective gradient in region to be matched and slope aspect and described the The one region to be matched characteristic point the most identical with described second region to be matched, as the cut-point of subregion, is treated described first Matching area and described second region to be matched are intactly divided into the subregion of several correspondence.

Characteristic point matching method towards up short stereoscopic image data the most according to claim 2, it is characterised in that: institute State and overlap between the first region to be matched or described second respective adjacent subarea territory, region to be matched.

Characteristic point matching method towards up short stereoscopic image data the most according to claim 2, it is characterised in that: step Suddenly the circulation of the described characteristic point in (5) forward coupling includes:

The first step, setting preset loop number of times and triangle size threshold value；

Second step, using each triangle in step (4) gained triangular network as first order triangle, with the described first order three Dihedral is constraints, utilizes described nearest neighbor distance algorithm to carry out the first order triangle by described first region to be matched to institute State the characteristic point forward coupling of the corresponding triangle in the second region to be matched, obtain the corresponding image points of each described first order triangle, N is set as 1, and m is set as 0；

3rd step, n=n+1, utilize the described cut-point conduct of the corresponding image points of correct coupling and subregion in the n-th 1 grades of trianglees Three summits of triangle build n-th grade of triangle；

4th step, with described n-th grade of triangle as constraints, utilize described nearest neighbor distance algorithm to carry out by described first and treat Described n-th grade of triangle of matching area mates to the characteristic point forward of the corresponding triangle in described second region to be matched, and leads to Cross repeatedly circulation RANSAC algorithm and reject error hiding characteristic point, draw the corresponding image points of correct coupling, m=m+1；

5th step, judge whether whether m be respectively less than described not less than the size of described preset loop number of times and n-th grade of triangle Triangle size threshold value, when m is respectively less than described three not less than the size of described preset loop number of times and described n-th grade of triangle During dihedral size threshold value, obtain forward corresponding image points group, otherwise return and perform the 3rd step,

Wherein, n is the rank of triangle, and m is cycle-index.

Characteristic point matching method towards up short stereoscopic image data the most according to claim 2, it is characterised in that: step Suddenly the reverse coupling of the circulation of the described characteristic point in (8) includes:

The first step, setting preset loop number of times and triangle size threshold value；

Second step, using each triangle in step (7) gained triangular network as first order triangle, with the described first order three Dihedral is constraints, utilizes described nearest neighbor distance algorithm to carry out the first order triangle by described second region to be matched to institute The characteristic point of the corresponding triangle stating the first region to be matched is inversely mated, and obtains the corresponding image points of each described first order triangle, N' is set as 1, and m' is set as 0；

3rd step, n'=n'+1', utilize the corresponding image points of correct coupling in the n-th ' 1 grade triangle to make with the described cut-point of subregion For three summits of triangle build n-th ' level triangle；

4th step, with described n-th ' level triangle as constraints, utilize described nearest neighbor distance algorithm to carry out by described second and treat Matching area described n-th ' level triangle inversely mates to the characteristic point of the corresponding triangle in described first region to be matched, and leads to Cross repeatedly circulation RANSAC algorithm and reject error hiding characteristic point, draw the corresponding image points of correct coupling, m'=m'+1；

5th step, judge m' whether not less than described preset loop number of times and n-th ' whether the size of level triangle be respectively less than described Triangle size threshold value, when m' not less than described preset loop number of times and described n-th ' the size of level triangle is respectively less than described three During dihedral size threshold value, obtain reverse corresponding image points group, otherwise return and perform the 3rd step,

Wherein, n' is the rank of triangle, and m' is cycle-index.