CN106228122A - planetary surface feature matching method based on set similarity - Google Patents

Abstract

Planetary surface feature matching method based on set similarity, the present invention relates to planetary surface feature matching method.The present invention is that the feature used to solve planet final landing section vision guided navigation lacks identification, in turn results in the problem that error hiding rate is higher.Feature Points Matching is converted to gather similarity and asks for problem by the present invention, it is proposed that a kind of visual signature matching algorithm being independent of image area information, and compares analysis with existing image local feature matching algorithm.For visual signature matching problem in planet image, the inventive method compares existing visual signature matching algorithm, as SURF describes son and template matching, while improving the characteristic matching rate of 10%, at least reduces the feature error hiding rate of 7%.The present invention is applied to space industry.

Description

Planetary surface feature matching method based on set similarity

Technical field

The present invention relates to planetary surface feature matching method.

Background technology

1, the difficulty that planetary landing task vision guided navigation is faced

Being in landing security consideration, landing point often selects at flat site, and its correspondence image texture is the most sparse, makes Become characteristics of image identification relatively low.Therefore, at lander when landing latter end is close to touchdown area, image local feature detection is calculated Although method is able to detect that a number of feature, but due near feature texture sparse, cause examined feature to lack identification, And then have a strong impact on the effect of characteristic matching, as shown in Figure 1.

In Fig. 1, picture is from the shooting over the ground of NEAR asteroid final landing.Wherein wire part is corresponding is declining image The location of pixels of middle Harris detection, the right is database images, and the characteristic point wherein detected from landing chart picture is counting Do not exist according in storehouse, but still be able to search match point (matching characteristic at three below in figure), in Fig. 1 by template matching Give three database feature point field image-regions that similarity ranking is higher, but at three, region is error hiding.It is not difficult Finding out, the identification of the periphery texture of three characteristic points is the most relatively low, presents high similarity.

2, domestic and international present Research

In planetary landing vision navigation system, generally require a property data base demarcating global position in advance, Land device declines process and shoots image over the ground, and extracts local feature in the picture, and obtains absolute pose with database matching Information.Conventional feature detection operator such as Harris angle point, SUSAN operator and SIFT scale invariant feature detection etc. by Maturation is applied at numerous areas.After feature detection completes, generally require refinement characterization information, and generate invariant with In data base, feature is mated.Lewis proposes template matching concept in nineteen ninety-five, utilizes the region of image normalization to be correlated with Property carry out matching characteristic, after this, some scholars template matching has been carried out improve and be applied in association area, as introduce combine Close discriminant function (Synthetic Discriminant Functions) and host element projection (Principal Components Projection) human eye template matching is carried out, three-dimensionalreconstruction based on extensive hierarchical mode template matching, and trunk three Dimension reconstruct etc..Essentially, the mathematical thought that template matching uses is vector similarity criterion, and this theory is in subsequent characteristics Allotment of labor has obtained further development in making.On the basis of scale invariant feature (SIFT) detection algorithm that Lowe is proposed in early stage, Proposed SIFT feature in 2004 and describe operator, by statistical nature point under characteristic dimension the gradient modulus value of near zone and Directional information, generates 128 dimensional features and describes son.This operator has been widely used in numerous areas, such as vision guided navigation, figure As inspection, three-dimensional body reconstruct etc..Additionally, some scholars have also carried out efficiency optimization to SIFT, Ke Yan et al. propose based on The PCA-SIFT of principal component analysis simplifies algorithm, 128 original dimension SIFT feature is described son and falls below 36 dimensions, ensure that essence Degree improves the computational efficiency of algorithm simultaneously.Bay uses box filtering to replace difference of Gaussian (DoG), it is proposed that Yi Zhonggeng Fast algorithm (Speeded Up Robust Features, SURF operator), compares SIFT algorithm, and calculating effect is greatly improved Rate.

Along with many optics secondary navigation systems constantly make progress in planetary exploration task, scholars are not in recent years Carry out vision additional planetary landing algorithm and be also carried out further investigation.Representative work is the JPL laboratory vision of NASA Navigation group, group member Cheng Yang et al. proposes to utilize crater as natural landmark in NEAR exploration task Carry out vision localization, and in research work subsequently, propose crater based on Canny edge detection operator quickly detect Algorithm, and propose to build radiation invariant to carry out crater coupling.AI.Mourikis et al. utilizes Harris angle point and template Matching algorithm (Template Matching), utilizes height and Attitude estimation information will detect characteristic point and data base spy of the same name Levy point scale difference to eliminate, add matching accuracy rate, and combine Mars accuracy mission requirements to vision aided inertial navigation Carry out systemic checking.

Summary of the invention

The present invention is that the feature used to solve planet final landing section vision guided navigation lacks identification, in turn results in The problem that error hiding rate is higher, and the planetary surface feature matching method based on set similarity proposed.

Planetary surface feature matching method based on set similarity realizes according to the following steps:

Step one: set in image the position of detected characteristic point as [μ_a,v_a], its correspondence same place in data base Picture position is [μ_D,v_D], [μ_a,v_a] and [μ_D,v_D] carry out affine transformation；

Step 2: three non-colinear characteristic point one trianglees of composition in image, its area is S_a；

Step 3: setting up the triangle of three characteristic points of the same name in data base, its area is S_D, S_aWith S_DBetween exist right Should be related to；

Step 4: for four characteristic points A, B, C, D, wherein the area ratio i.e. structure of triangle Δ ACD and triangle Δ ABD Become affine invarient f_inv；

Step 5: build matrix I_invRepresent the affine invarient of each Feature point correspondence in image；

Step 6: obtain f according to step 4 and step 5_inv；

Step 7: according to the matrix I in step 5_invObtainJaccard is used to gather similarity criterion；

Step 8: set up the coupling between affine invarient, matching criteria is d_ij；

Step 9: set up complete set similarity and ask for algorithm.

Invention effect:

Existing Feature Correspondence Algorithm is all set up on the basis of to feature detection periphery texture information, but appoints in planetary landing Business latter end, in image, the feature texture of low identification can significantly improve the error hiding rate of feature, and then causes vision guided navigation sternly Ghost image rings, it is necessary to take into account design new Feature Correspondence Algorithm to reduce error hiding rate.Considering factors above, the present invention is by feature Point matching is converted to gather similarity and asks for problem, it is proposed that a kind of visual signature coupling being independent of image area information is calculated Method, and compare analysis with classical image local feature matching algorithm.

For visual signature matching problem in planet image, the carried algorithm of the present invention is compared existing visual signature coupling and is calculated Method (as SURF describes son, template matching), while improving the characteristic matching rate of 10%, at least reduces the feature mistake of 7% Join rate.

Accompanying drawing explanation

Fig. 1 is Harris detection and template matching exemplary plot；

Fig. 2 is affine invarient schematic diagram based on triangle area ratio；

Fig. 3 is data base and decline image simulation simulated example figure；

Fig. 4 is the exemplary plot of affine ratio iteration of variables matching process.

Detailed description of the invention

Detailed description of the invention one: planetary surface feature matching method based on set similarity comprises the following steps:

Step one: consider that the identification of landing latter end image local feature is relatively low, herein propose the image utilizing characteristic point Position builds the invariant for characteristic matching.If the position of detected characteristic point is [μ in image_a,v_a], its correspondence is in data The picture position of the same place in storehouse is [μ_D,v_D], because characteristic point is seated in the landing point field of relatively flat, therefore can be false If characteristic point is coplanar in data base, [μ_a,v_a] and [μ_D,v_D] carry out affine transformation；

Step 2: three non-colinear characteristic point one trianglees of composition in image, its area is S_a；

Step 3: setting up the triangle of three characteristic points of the same name in data base, its area is S_D, S_aWith S_DBetween exist right Should be related to；

Step 4: as in figure 2 it is shown, for four characteristic points A, B, C, D, wherein triangle Δ ACD and triangle Δ ABD Area ratio i.e. constitutes affine invarient f_inv；

Step 5: build matrix I_invRepresent the affine invarient of each Feature point correspondence in image；

Step 6: obtain f according to step 4 and step 5_inv；

Step 7: according to the matrix I in step 5_invObtainJaccard is used to gather similarity criterion；

Step 8: set up the coupling between affine invarient, matching criteria is d_ij；

Step 9: set up complete set similarity and ask for algorithm.

Detailed description of the invention two: present embodiment is unlike detailed description of the invention one: the S in described step 2_aTool Body is:

Wherein, H_DFor homography matrix, T=[t_x,t_y]^TIt it is the relative translation vector of two width images.

Other step and parameter are identical with detailed description of the invention one.

Detailed description of the invention three: present embodiment is unlike detailed description of the invention one or two: in described step 2 S_aParticularly as follows:

$S_a=\frac{1}{2}\det \left[\begin{array}{ccc}{\mathrm{\μ}}_1& {\mathrm{\ν}}_1& 1\\ {\mathrm{\μ}}_2& {\mathrm{\ν}}_2& 1\\ {\mathrm{\μ}}_3& {\mathrm{\ν}}_3& 1\end{array}\right]---\left(2\right)$

Wherein μ₁、μ₂、μ₃、v₁、v₂、v₃It is respectively the location of pixels of three characteristic points in image.

Other step and parameter are identical with detailed description of the invention one or two.

Detailed description of the invention four: present embodiment is unlike one of detailed description of the invention one to three: described step 3 Middle S_aWith S_DBetween there is corresponding relation particularly as follows:

S_D=| det (H) | S_a (3)

Other step and parameter are identical with one of detailed description of the invention one to three.

Detailed description of the invention five: present embodiment is unlike one of detailed description of the invention one to four: described step 4 Middle affine invarient f_invParticularly as follows:

$f_{inv}=\left|\det \left(H\right)\right|=\frac{S_D\left(\mathrm{\Δ}ABD\right)}{S\left(\mathrm{\Δ}ABD\right)}=\frac{S_D\left(\mathrm{\Δ}ACD\right)}{S\left(\mathrm{\Δ}ACD\right)}---\left(4\right)$

Other step and parameter are identical with one of detailed description of the invention one to four.

Detailed description of the invention six: present embodiment is unlike one of detailed description of the invention one to five: described step 5 Middle matrix I_invParticularly as follows:

Fig. 2 declines each characteristic point in image and is all attached to 3 independent trianglees, therefore can build individual affine constant Amount, builds following matrix and describes the affine invarient of each Feature point correspondence in image at this:

It is seen that, if image I exists n characteristic point, then can corresponding existIndividual attached triangle, it is affine not The size of bending moment battle array is then

Other step and parameter are identical with one of detailed description of the invention one to five.

Detailed description of the invention seven: present embodiment is unlike one of detailed description of the invention one to six: described step 6 In f_invParticularly as follows:

Due to f_invAffine invarient structure in, the division of triangle has secondary sequence, and in characteristic matching, should Order is unpredictable, therefore by affine invarient f_invChange as follows, obtain f_inv:

$f_{inv}=\frac{S\left({\mathrm{\Δ}}_1\right)}{S\left({\mathrm{\Δ}}_2\right)}+\frac{S\left({\mathrm{\Δ}}_2\right)}{S\left({\mathrm{\Δ}}_1\right)}---\left(6\right)$

Wherein Δ₁,Δ₂The area of any two triangle in separated image, such as Δ₁=S (Δ ABC), Δ₂=S (Δ ABD), Δ is worked as₁,Δ₂Division reversed order time, new invariant f_invNo longer change.

Other step and parameter are identical with one of detailed description of the invention one to six.

Detailed description of the invention eight: present embodiment is unlike one of detailed description of the invention one to seven: described step 7 Middle according to the matrix I in step 5_invObtainThe detailed process using Jaccard to gather similarity criterion is:

According to matrix I_inv, arbitrary characteristics point [μ in image_i,v_i]^T, i=1 ..., the affine invarient set description of n is:

WhereinIt is the i-th row of formula (5), does not possess order owing to the affine invarient of characteristic point builds, thereforeIt is more suitable for describing with the form of set；Therefore in formula (5), the form of affine constant matrix is rewritten as:

So far, the task of Feature Points Matching just can be converted into the similarity asked between two set, uses Jaccard collection Close similarity criterion, be shown below:

WhereinIt is the set of two affine invarient, wherein | | represent the element number in set (Cardinality, also known as radix).

Other step and parameter are identical with one of detailed description of the invention one to seven.

Detailed description of the invention nine: present embodiment is unlike one of detailed description of the invention one to eight: described step 8 In d_ijParticularly as follows:

Above formula is asked for it is to be appreciated that gatherThe radix of common factor, it is therefore desirable to set up between affine invarient Join, herein propose the matching criteria of a kind of improvement:

$d_{ij}=|1-\exp (-|2-f_i^h/f_j^D-f_j^D/f_i^h\left|\right)|---\left(10\right)$

Wherein f_iIt is to decline the i-th affine invarient of feature h in image,It is that in data base, the jth of feature D is affine Invariant；Work as d_ijDuring less than setting threshold value, affine invarientSuccessful match.

Detailed description of the invention ten: present embodiment is unlike one of detailed description of the invention one to nine: described step 9 The complete set similarity of middle foundation is asked for algorithm detailed process and is:

Input is:For declining the affine invarient set of characteristics of image h,For the affine invarient set of kth feature, S in data base_d[k]=0, C_d[k]=0, Distance threshold: T_d；

Set up affine invarient Distance Judgment matrix:Wherein d_ij=D (i, j) correspondingI-th element withThe distance of jth element, from beeline d_minStart square Battle array D is addressed that sequence is until distance d > T_dTill (as shown in Figure 4, least member in iterative matrix, and delete correspondence Row and column, until least member d in matrix_min>T_d, record solves number of times C_d[k]=i_nWith least member sumAnd record the element number C meeting distance criterion_d[k], thenWithSet similarity For:

Wherein saidIn representation database, feature k penetrates invariant set；Represent and decline feature h in image Affine invarient set, T_dFor affine invarient matching threshold；

Behind ergodic data storehouse, feature h will select J_maxCorresponding feature is as matching result；If existing and solving situation, then more Select total distance S_dMinimum feature is as final matching results.

Other step and parameter are identical with one of detailed description of the invention one to nine.

Embodiment one:

For real-time and the effectiveness of testing algorithm, selection planet true picture is as test sample, first in high-resolution The planet image of degree carries out the subregion in feature point detection, and random cropping planet image carry out affine transformation and simulate Different during decline camera shooting visual angles, highly.

As it is shown on figure 3, the left side is database images in figure, the right side 6 width subimage is the parallel affine transformation of random cropping Decline stage analog image.First by SURF, database images is carried out feature detection, on this basis, choosing with declining image Select SURF Feature Descriptor matching algorithm and the template matching algorithm group as a comparison of classics.By 30 width planets are truly schemed As carrying out emulation experiment, obtain following statistical result:

Table 1 Image Feature Matching results contrast

As shown above, four kinds of algorithms are emulated by the desktop computer using 2.4Ghz, 8G internal memory herein.Wherein four kinds Algorithm employs identical SURF feature detection, for comparing database feature, average detected rate and false detection rate be divided into for 90% and 5%.In simulations, average each image contains 7.1 characteristic points.AISM algorithm is the most excellent in matching rate and error hiding rate In SURF Feature Descriptor and template matching algorithm, wherein real-time and template matching algorithm are more or less the same but are substantially better than SURF Feature Correspondence Algorithm.

Claims (10)

1. planetary surface feature matching method based on set similarity, it is characterised in that described planetary surface characteristic matching side Method comprises the following steps:

Step one: set in image the position of detected characteristic point as [μ_a,ν_a], the image of its correspondence same place in data base Position is [μ_D,ν_D], [μ_a,ν_a] and [μ_D,ν_D] carry out affine transformation；

Step 2: three non-colinear characteristic point one trianglees of composition in image, its area is S_a；

Step 3: setting up the triangle of three characteristic points of the same name in data base, its area is S_D, S_aWith S_DBetween exist correspondence close System；

Step 4: for four characteristic points A, B, C, D, wherein triangle △ ACD i.e. constitutes with the area ratio of triangle △ ABD Affine invarient f_inv；

Step 5: build matrix I_invRepresent the affine invarient of each Feature point correspondence in image；

Step 6: obtain f according to step 4 and step 5_inv；

Step 7: according to the matrix I in step 5_invObtainJaccard is used to gather similarity criterion；

Step 8: set up the coupling between affine invarient, matching criteria is d_ij；

Step 9: set up complete set similarity and ask for algorithm.

Planetary surface feature matching method based on set similarity the most according to claim 1, it is characterised in that described S in step 2_aParticularly as follows:

Wherein, H_DFor homography matrix, T=[t_x,t_y]^TIt it is the relative translation vector of two width images.

Planetary surface feature matching method based on set similarity the most according to claim 2, it is characterised in that described S in step 2_aParticularly as follows:

$S_a=\frac{1}{2}\det \left[\begin{array}{ccc}{\mathrm{\μ}}_1& v_1& 1\\ {\mathrm{\μ}}_2& v_2& 1\\ {\mathrm{\μ}}_3& v_3& 1\end{array}\right]---\left(2\right)$

Wherein μ₁、μ₂、μ₃、ν₁、ν₂、ν₃It is respectively the location of pixels of three characteristic points in image.

Planetary surface feature matching method based on set similarity the most according to claim 3, it is characterised in that described S in step 3_aWith S_DBetween there is corresponding relation particularly as follows:

S_D=| det (H) | S_a (3)。

Planetary surface feature matching method based on set similarity the most according to claim 4, it is characterised in that described Affine invarient f in step 4_invParticularly as follows:

$f_{inv}=\left|\det \left(H\right)\right|=\frac{S_D\left(\mathrm{\Δ}ABD\right)}{S\left(\mathrm{\Δ}ABD\right)}=\frac{S_D\left(\mathrm{\Δ}ACD\right)}{S\left(\mathrm{\Δ}ACD\right)}---\left(4\right).$

Planetary surface feature matching method based on set similarity the most according to claim 5, it is characterised in that described Matrix I in step 5_invParticularly as follows:

Planetary surface feature matching method based on set similarity the most according to claim 6, it is characterised in that described F in step 6_invParticularly as follows:

$f_{inv}=\frac{S\left({\mathrm{\Δ}}_1\right)}{S\left({\mathrm{\Δ}}_2\right)}+\frac{S\left({\mathrm{\Δ}}_2\right)}{S\left({\mathrm{\Δ}}_1\right)}---\left(6\right)$

Wherein △₁,△₂In separated image, the area of any two triangle, works as △₁,△₂Division reversed order time, f_invNo longer Change.

Planetary surface feature matching method based on set similarity the most according to claim 7, it is characterised in that described According to the matrix I in step 5 in step 7_invObtainThe detailed process using Jaccard to gather similarity criterion is:

According to matrix I_inv, arbitrary characteristics point [μ in image_i,ν_i]^T, i=1 ..., the affine invarient set description of n is:

WhereinIt is the i-th row of formula (5), does not possess order owing to the affine invarient of characteristic point builds, therefore willDescribe with the form of set；Therefore in formula (5), the form of affine constant matrix is:

Feature Points Matching is converted into the similarity asked between two set, uses Jaccard to gather similarity criterion, such as following formula institute Show:

WhereinIt it is the set of two affine invarient.

Planetary surface feature matching method based on set similarity the most according to claim 8, it is characterised in that described D in step 8_ijParticularly as follows:

$d_{ij}=|1-\exp (-|2-f_i^h/f_j^D-f_j^D/f_i^h\left|\right)|---\left(10\right)$

Wherein f_iIt is to decline the i-th affine invarient of feature h in image,It is that in data base, the jth of feature D is affine constant Amount；Work as d_ijDuring less than setting threshold value, affine invarientSuccessful match.

Planetary surface feature matching method based on set similarity the most according to claim 9, it is characterised in that institute State step 9 is set up complete set similarity to ask for algorithm detailed process and be:

Set up affine invarient Distance Judgment matrix: D_m×n=[..., d_ij...], wherein d_ij=D (i, j) correspondingI-th Individual element withThe distance of jth element, from beeline d_minStart matrix D is addressed sequence until distance d >T_dTill, and record the element number C meeting distance criterion_d[k], thenWithSet similarity be:

Wherein saidIn representation database, feature k penetrates invariant set；In representative decline image, feature h is imitative Penetrate invariant set, T_dFor affine invarient matching threshold；

Behind ergodic data storehouse, feature h will select J_maxCorresponding feature is as matching result；If existing and solving situation more, then select Total distance S_dMinimum feature is as final matching results.