CN109389625A - A kind of three-dimensional image registration method screening out error hiding based on multiple dimensioned description - Google Patents

Abstract

The invention discloses a kind of three-dimensional image registration method for screening out error hiding based on multiple dimensioned description, this method preferably describes the feature of corresponding key point, and tentatively obtain Corresponding matching point by constructing a kind of novel multiple dimensioned description；And traverse as feature has the point of similar multiple dimensioned description to converge therewith in subject to registration cloud, greatly improve an operational efficiency for cloud rough registration, reduce the calculation amount of computer, bring great convenience for point cloud registering；This method can obtain more accurate registration effect in the shorter time, and robustness is more preferable, suitable for there are noise, structure is complicated, to being registrated demanding field of precision measurement.

Description

A kind of three-dimensional image registration method screening out error hiding based on multiple dimensioned description

Technical field

The invention belongs to image registration field, in particular to a kind of three-dimensional figure that error hiding is screened out based on multiple dimensioned description As method for registering.

Background technique

Demand with electronics, the rapid development of computer technology and people to high quality of products, external machine vision Enter high-speed development period in the 1990s, and is widely used in industrial control field.Machine vision is to realize that industry is automatic Change and intelligentized necessary means, are equivalent to extension of the human vision on machine.Machine vision has increasingly automated, efficient The advantages that rate, high-precision, will also play great effect in China's automatic field.It is regarded compared to traditional two-dimentional machine Feel, the precision of three-dimensional machine vision is higher, detection effect is more preferable, is able to satisfy harsher production technology.But three-dimensional machine The point cloud data amount that vision obtains is big and between points without apparent topological relation, therefore we are during registration It is easy to appear and mismatches quasi- right, this will be such that our registration accuracy reduces, and influence measurement and detection effect to object.

In aerospace, the intelligence manufactures technique such as mechanical transport, bridge ship, adding for Complex Different Shape curved surface can be all encountered Work and detection, curved surface complex bend is simultaneously folded in space arbitrarily product, and edge is complicated, and two-dimentional machine vision is difficult to meet its height Precision, efficient production and processing requirement.But existing point cloud registration algorithm accuracy rate is not high, largely mismatches quasi- to depositing In the precision and efficiency for affecting registration, detection and reconstruction to object, which also result in, to be seriously affected.

It is directed to this problem at present, most widely used solution is RANSAC algorithm, although RANSAC algorithm is big It works well in most cases, but its runing time can be exponentially increased with the ratio η of abnormal point pair, therefore the calculation Method is very sensitive to the selection of sample point in point cloud registering, and robustness is poor.During registration, the distribution of point cloud density is Irregular and its structure also lacks regularity, so that the accuracy rate of 3-D image registration is more much lower than two dimensional image.

Summary of the invention

The present invention provides a kind of three-dimensional image registration method that error hiding is screened out based on multiple dimensioned description, this method benefits It is screened with the space geometry relationship between registration point pair, and then can directly reject and mismatch a little, improve registration efficiency and match Quasi- precision.

A kind of three-dimensional image registration method screening out error hiding based on multiple dimensioned description, comprising the following steps:

Step 1: obtaining the three-dimensional point cloud X and Y of the adjacent scene in visual angle, and more for each point construction in three-dimensional point cloud X and Y Scale describes son (Q_i, N_i)；

Wherein, l indicates the neighborhood search radius label at the midpoint three-dimensional point cloud X or Y, and value is followed successively by 1,2 ..., and L, L are indicated Neighborhood search radius sum, L >=2；T_lIndicate normalized vector,λ_l1, λ_l2, λ_l3For Matrix C_lIt carries out Characteristic value after SVD decomposition, λ_l1≥λ_l2≥λ_l3, n_l1, n_l2, n_l3Respectively λ_l1, λ_l2, λ_l3Corresponding feature vector,S_l={ p_i|||p_i-p₀||≤r_l, | S_l| indicate S_lMaximum value, p_iIndicate three Certain point p in dimension point cloud X or Y₀Neighborhood point, r_lIndicate point p₀First of neighborhood search radius,r_LFor setting Maximum neighborhood search radius, n_lRespectively indicate the point normal vector corresponding with the search radius of neighbourhood l of point in three-dimensional point cloud, value For n_l3；

|S_l| it is with p₀For the centre of sphere, r_lFor the maximum value in the distance of all the points in the ball of the centre of sphere to the centre of sphere；

Step 2: multiple dimensioned description of each point in traversal three-dimensional point cloud X and Y will have identical multiple dimensioned retouch in X and Y The point for stating son is recorded as crucial point set A respectively_i0And B_i0；

Step 3: according to principal curvatures signature search key point set A_i0And B_i0In initial matching point set A_i1And B_i1；

Initial matching point set A_i1In point a_i1And B_i1In b_i1It is matching double points；

From crucial point set A_i0Choose arbitrary point a_i, traverse B_i0Midpoint is selected and point a_iThe point b to match_j, following to meet The point of condition constructs initial matching point set A_i1And B_i1:

Wherein, k₁(a_i) and k₂(a_i) it is illustrated respectively in point a_iPlace, utilizes point a_iNeighborhood point obtain principal curvatures minimum With principal curvatures maximum, k₁(b_j) and k₂(b_j) it is illustrated respectively in point b_jPlace, utilizes point b_jNeighborhood point obtain principal curvatures it is minimum Value and principal curvatures maximum, δ₁For the first error threshold value of setting；

Step 4: using similar triangles threshold value to initial matching point set A_i1And B_i1In matching double points screened, obtain Registration point is obtained to collection A_i2And B_i2；

Step 5: A is calculated based on SVD_i2And B_i2Between the first spin matrix R₁With the first translation matrix T₁, utilize One spin matrix R₁With the first translation matrix T₁, successively calculate A_i2In each point a_i2Change point a_3iCoordinate, while calculating A_i2 In each point in B_i2In match point b_2iWith the Euclidean distance d between correspondent transform point_3i；

A_i3=R₁*A_i2+T₁

Wherein, (a_3ix,a_3iy,a_3iz) and (b_i2x,b_i2y,b_i2z) it is respectively a_3iAnd b_i2Coordinate on x, y, z axis；

Step 6: calculatingThe initial value of k is that 1, N is current registration point to collection A_i2And B_i2Middle registration point Logarithm, and judge d_k+1-d_k< ξ₃It is whether true, if so, then with current registration point to collection A_i2And B_i2As three-dimensional point cloud X and Y Registration data otherwise enable k=k+1, return step 3 chooses each neighborhood of a point point again, calculates principal curvatures.

Further, rotation angle threshold θ is utilized_mRegistration is screened again to doing, by current registration point to collection A_i2And B_i2 In, it is unsatisfactory for the corresponding registration point of following formula and is rejected to as abnormal point or noise spot:

Δθ_i< θ_m

Wherein, Δ θ_iIndicate current registration point to collection A_i2And B_i2Rotation angle difference between middle i-th pair registration point pair；

Current registration point is to collection A_i2And B_i2Middle i-th pair registration point is to (x_i,y_i) between rotation angle difference solution Journey is as follows:

1) from current registration point to collection A_i2And B_i2In, non-coplanar two pairs of registrations pair are randomly selected, it is non-coplanar with four Point constitutes a spherical surface, finds out affiliated spherical equation and centre of sphere O by analytic geometry and establishes space coordinates using point O as origin；

2) using the Z axis for the space coordinates built as the second spin matrix R₂Rotary shaft, and R₂=A₂B₂, A₂And B₂Point The second spin matrix R is not indicated₂Split-matrix, B₂By current registration point to collection A_i2And B_i2Middle a pair of registration point is to (x_k,y_k) Meet B₂x_k=y_kWhen, carry out calculating acquisition, A₂Rotary shaft be Z axis, meanwhile, according to the registration distance threshold ξ of setting, calculate It is registrated angle threshold ε_i,

3) rotation byThe border circular areas surrounded, whenThe left margin of the border circular areas surrounded with y_iWhen tangent, byThe rotation angle of the border circular areas surrounded is θ₁, at this time byThe circle surrounded The center of circle in domain is point M, by point M and y_iDistance r_sA quadratic equation with one unknown is constructed, using equation of the point M in great circle O, Two equation groups of simultaneous obtain the x and y coordinates value of point M, find out θ₁；

Indicate point x_iBy B₂Rigid body translation is formed with B₂x_iFor the centre of sphere, ε_iFor the border circular areas of radius；

Because of B₂It is the rigid body translation spin matrix when having noise or abnormal point,It indicates: x_iThrough Cross B₂Rigid body translation, but because there are noise or abnormal point, can only find out a satisfactory range, and this range It is exactly with B₂x_iFor the center of circle, ε_iFor the border circular areas of radius；

The great circle O is to give directions x_iIn shown space coordinates, using point O as the circle where origin；

4) continue to rotateThe border circular areas surrounded, byThe right margin of the border circular areas surrounded With y_iWhen tangent,The rotation angle of the border circular areas surrounded is θ₂, at this time byThe circle surrounded The center of circle in domain is point N, by point N and y_iDistance r_sA quadratic equation with one unknown is constructed, equation of the N in great circle O, connection are utilized Vertical two equation groups obtain the x and y coordinates value of point N, find out θ₂, obtain angle difference Δ θ_i；

Wherein, r_sIt isThe radius of the border circular areas surrounded, г indicate the B in space coordinates O₂x_iTo point O Distance, r_s=г * tan ε_i。

Further, the principal curvatures calculating process of each point is as follows in the three-dimensional point cloud:

Step 3.1: in three-dimensional point cloud arbitrary point and the neighborhood of a point point be fitted to a local secondary curved surface G (x, y)；

G (x, y)=ax²+bxy+cy²+dx+ey

Step 3.2: derivation being carried out to local quadratic surface G (x, y), it is basic to obtain first fundamental constant E, F, G and second Constant L, M, N；

E=G_xx, F=G_x*G_y, G=G_yy；

L=nG_xx, M=nG_xy, N=nG_yy；

Wherein, G_x、G_yRespectively G (x, y) asks x, y and once leads, G_xx、G_yyIt asks secondary to x, y for G (x, y) to lead, G_xyFor G (x, y) successively to x and y derivation, n is normal vector of the local secondary curved surface G (x, y) at selected arbitrary point；

Step 3.3: calculate Gaussian curvature K and mean curvature H:

Step 3.4: calculating principal curvatures minimum k₁With principal curvatures maximum k₂:

Further, described to utilize similar triangles threshold value to initial matching point set A_i1And B_i1In matching double points carry out The process of screening is as follows:

Step 4.1: in A_i1In randomly select 3 not conllinear point a_i,a_j,a_k, in B_i1Find corresponding 3 point b_u,b_v, b_w；

Step 4.2: enabling (a_i,b_u) and (a_j,b_v) it is known matching double points, it is carried out according to Similar Principle of Triangle following It calculates:

Δa_ij=| | a_i-a_j||；Δa_ik=| | a_i-a_k||；Δa_jk=| | a_j-a_k||；

Δb_uv=| | b_u-b_v||；Δb_uw=| | b_u-b_w||；Δb_wv=| | b_w-b_v||；

Step 4.3: by Similar Principle of Triangle, calculating two triangle similarity Ω；

Wherein, Δ a_ij,Δa_ik,Δa_jk；Δb_uv,Δb_uw,Δb_wvFor each side length of triangle, δ₂It is the second of setting Error threshold；

Step 4.4: if a_i,a_j,a_k, b_u,b_v,b_wMeet Ω < δ₂, then it is assumed that a_kAnd b_wBelong to a pair of of registration point pair.

Further, the L value is 4, and r₄=0.004m.

Further, the δ₂Value is any one in 0.02,0.05,0.1.

Further, the three-dimensional point cloud of the adjacent scene in visual angle is obtained using 3D profile scanning sensor.

Beneficial effect

The present invention provides a kind of three-dimensional image registration methods that error hiding is screened out based on multiple dimensioned description, pass through construction A kind of novel multiple dimensioned description, preferably describes the feature of corresponding key point, and tentatively obtain Corresponding matching point；And with this Be characterized in subject to registration cloud of traversal has the point of similar multiple dimensioned description to converge therewith, greatly improves a cloud rough registration Operational efficiency reduces the calculation amount of computer, brings great convenience for point cloud registering；This method can be when shorter Between in obtain more accurate registration effect, and robustness is more preferable, suitable for there are noise, structure is complicated, to high with alignment request Field of precision measurement.

After essence registration, the point cloud pair being registrated can inevitably exist compared to traditional method for registering and mismatch standard Pair and noise, accuracy rate and precision of the objective in measurement and detection certainly will be will affect.Application now is at most It is ICP iterative method to carry out optimal registration, but largely mismatches quasi- pair of presence, largely will increase algorithm and fall into part A possibility that optimal or algorithm can not restrain seriously affects the reliability of registration.Therefore, the present invention is based on rotation angle by introducing Degree threshold value mismatches quasi- right further to screen out, so that the quantity for mismatching quasi- pair and noise spot largely reduces, ensure that registration knot The accuracy of fruit makes it have good robustness；On the other hand, quasi- pair and noise spot are mismatched due to having screened out a part, The input quantity of a cloud is decreased, the calculation amount in registration process is reduced, improves the efficiency of registration.

Detailed description of the invention

Fig. 1 is the flow diagram of the method for the invention；

Fig. 2 is that of the present invention utilize rotates angle progress registration process schematic diagram, wherein figure (a) is space coordinate It is schematic diagram, figure (b) is that threshold xi is converted into angle ε_iSchematic diagram, figure (c) be x_iAnd x_kRespectively in R and R₂Signal under transformation Figure, figure (d) are B₂x_iRotary course schematic diagram, figure (e) beThe border circular areas surrounded is in space coordinates Relation schematic diagram.

Specific embodiment

Below in conjunction with drawings and examples, the present invention is described further.

As shown in Figure 1, a kind of three-dimensional image registration method that error hiding is screened out based on multiple dimensioned description, including following step It is rapid:

Step 1: obtaining the three-dimensional point cloud X and Y of the adjacent scene in visual angle, and more for each point construction in three-dimensional point cloud X and Y Scale describes son (Q_i, N_i)；

The three-dimensional point cloud of the adjacent scene in visual angle is obtained in this example using 3D profile scanning sensor；

Wherein, l indicates the neighborhood search radius label at the midpoint three-dimensional point cloud X or Y, and value is followed successively by 1,2 ..., and L, L are indicated Neighborhood search radius sum, L >=2；T_lIndicate normalized vector,λ_l1, λ_l2, λ_l3For Matrix C_lIt carries out Characteristic value after SVD decomposition, λ_l1≥λ_l2≥λ_l3, n_l1, n_l2, n_l3Respectively λ_l1, λ_l2, λ_l3Corresponding feature vector,S_l={ p_i|||p_i-p₀||≤r_l, | S_l| indicate S_lMaximum value, p_iIndicate three Certain point p in dimension point cloud X or Y₀Neighborhood point, r_lIndicate point p₀First of neighborhood search radius,r_LFor setting Maximum neighborhood search radius, n_lRespectively indicate the point normal vector corresponding with the search radius of neighbourhood l of point in three-dimensional point cloud, value For n_l3；

In this example, the L value is 4, and r₄=0.004m.

|S_l| it is with p₀For the centre of sphere, r_lFor the maximum value in the distance of all the points in the ball of the centre of sphere to the centre of sphere；

Step 2: multiple dimensioned description of each point in traversal three-dimensional point cloud X and Y will have identical multiple dimensioned retouch in X and Y The point for stating son is recorded as crucial point set A respectively_i0And B_i0；

Step 3: according to principal curvatures signature search key point set A_i0And B_i0In initial matching point set A_i1And B_i1；

Initial matching point set A_i1In point a_i1And B_i1In b_i1It is matching double points；

From crucial point set A_i0Choose arbitrary point a_i, traverse B_i0Midpoint is selected and point a_iThe point b to match_j, following to meet The point of condition constructs initial matching point set A_i1And B_i1:

Wherein, k₁(a_i) and k₂(a_i) it is illustrated respectively in point a_iPlace, utilizes point a_iNeighborhood point obtain principal curvatures minimum With principal curvatures maximum, k₁(b_j) and k₂(b_j) it is illustrated respectively in point b_jPlace, utilizes point b_jNeighborhood point obtain principal curvatures it is minimum Value and principal curvatures maximum, δ₁For the first error threshold value of setting；

First error threshold value δ₁Size one best value of effect is acquired by experiment simulation；

The principal curvatures calculating process of each point is as follows in the three-dimensional point cloud:

Step 3.1: in three-dimensional point cloud arbitrary point and the neighborhood of a point point be fitted to a local secondary curved surface G (x, y)；

G (x, y)=ax²+bxy+cy²+dx+ey

Step 3.2: derivation being carried out to local quadratic surface G (x, y), it is basic to obtain first fundamental constant E, F, G and second Constant L, M, N；

E=G_xx, F=G_x*G_y, G=G_yy；

L=nG_xx, M=nG_xy, N=nG_yy；

Wherein, G_x、G_yRespectively G (x, y) asks x, y and once leads, G_xx、G_yyIt asks secondary to x, y for G (x, y) to lead, G_xyFor G (x, y) successively to x and y derivation, n is normal vector of the local secondary curved surface G (x, y) at selected arbitrary point；

Step 3.3: calculate Gaussian curvature K and mean curvature H:

Step 3.4: calculating principal curvatures minimum k₁With principal curvatures maximum k₂:

Step 4: using similar triangles threshold value to initial matching point set A_i1And B_i1In matching double points screened, obtain Registration point is obtained to collection A_i2And B_i2；

It is described to utilize similar triangles threshold value to initial matching point set A_i1And B_i1In the process screened of matching double points It is as follows:

Step 4.1: in A_i1In randomly select 3 not conllinear point a_i,a_j,a_k, in B_i1Find corresponding 3 point b_u,b_v, b_w；

Step 4.2: enabling (a_i,a_u) and (a_j,a_v) it is known matching double points, it is carried out according to Similar Principle of Triangle following It calculates:

Δa_ij=| | a_i-a_j||；Δa_ik=| | a_i-a_k||；Δa_jk=| | a_j-a_k||；

Δb_uv=| | b_u-b_v||；Δb_uw=| | b_u-b_w||；Δb_wv=| | b_w-b_v||；

Step 4.3: by Similar Principle of Triangle, calculating two triangle similarity Ω；

Wherein, Δ a_ij,Δa_ik,Δa_jk；Δb_uv,Δb_uw,Δb_wvFor each side length of triangle, δ₂It is the second of setting Error threshold；

In this example, the δ₂Value is any one in 0.02,0.05,0.1.

Step 4.4: if a_i,a_j,a_k, b_u,b_v,b_wMeet Ω < δ₂, then it is assumed that a_kAnd b_wBelong to a pair of of registration point pair.

Step 5: A is calculated based on SVD_i2And B_i2Between the first spin matrix R₁With the first translation matrix T₁, utilize One spin matrix R₁With the first translation matrix T₁, successively calculate A_i2In each point a_i2Change point a_3iCoordinate, while calculating A_i2 In each point in B_i2In match point b_2iWith the Euclidean distance d between correspondent transform point_3i；

A_i3=R₁*A_i2+T₁

Wherein, (a_3ix,a_3iy,a_3iz) and (b_i2x,b_i2y,b_i2z) it is respectively a_3iAnd b_i2Coordinate on x, y, z axis；

Step 6: calculatingThe initial value of k is that 1, N is current registration point to collection A_i2And B_i2Middle registration point Logarithm, and judge d_k+1-d_k< ξ₃It is whether true, if so, then with current registration point to collection A_i2And B_i2As three-dimensional point cloud X and Y Registration data otherwise enable k=k+1, return step 3 chooses each neighborhood of a point point again, calculates principal curvatures.

In this example, ξ₃Value is 0.01.

Utilize rotation angle threshold θ_mRegistration is screened again to doing, by current registration point to collection A_i2And B_i2In, it is unsatisfactory for The corresponding registration point of following formula is rejected to as abnormal point or noise spot:

Δθ_i< θ_m

Wherein, Δ θ_iIndicate current registration point to collection A_i2And B_i2Rotation angle difference between middle i-th pair registration point pair；

Current registration point is to collection A_i2And B_i2Middle i-th pair registration point is to (x_i,y_i) between rotation angle difference solution Journey is as follows:

1) from current registration point to collection A_i2And B_i2In, non-coplanar two pairs of registrations pair are randomly selected, it is non-coplanar with four Point constitutes a spherical surface, finds out affiliated spherical equation and centre of sphere O by analytic geometry and establishes space coordinates using point O as origin, As shown in Fig. 2 (a), and from Fig. 2 (a), it can be seen that with y_iThe public area intersected is had with sphere O for the sphere of the centre of sphere Domain, this region are considered y_iMeet the region of registration condition；

As long as x_iBy R₁、T₁Transformed point belong to y_iFor the centre of sphere, ξ is in the sphere of radius, then it is assumed that x_iAnd y_i It is a pair of of registration pair, i.e., the satisfaction to threshold xi is converted into angle ε_iSolution, as shown in Fig. 2 (b)；

Consider be based on a screening conditions (rotation transformation), to meet rotation and translation transformation when angle threshold value, The threshold value of rotation angle when must first meet based on rotation transformation, most homogeneous problem is converted are as follows: ∠ (R₁x_i, y_i)≤ε_i, wherein ε_iFor the angle threshold for judging whether it is registration point pair；

By Fig. 2 (b) it follows that

Based on above-mentioned analysis, it does such as down conversion:

1. R₁It is decomposed into two matrix As, B, i.e. R₁=A*B；(B is set to meet condition: ∠ (Bx_i,y_i)≤ε_i；Make A simultaneously For Bx_iRotary shaft, most homogeneous problem is converted into the rotation angle asked and A is made to meet condition as far as possible；

2. when no noise and abnormal point, x_iAnd y_iIt can perfectly be stitched together very much.Define one R in the ideal situation₂=A₂B₂.By a pair of registration to x_k,y_k, can be in the hope of B₂(B₂x_k=y_k)；Using the point of i ≠ k, with B₂x_kFor rotary shaft (i.e. A in Fig. 2 (c)₂Rotary shaft), find out it and make B₂x_iWith y_iThe rotation angle θ that perfection is mixed₂(because of point B₂x_iAnd y_iCoordinate in space coordinates is known, A₂Rotary shaft can be regarded as the Z axis in space coordinates, by The relationship of analytic geometry can directly find out rotation angle θ₂, that is, acquire A₂:

A₂=exp (θ₂[B₂x_k]_X)

Wherein, [X]_XIndicate the vector product of X, exp () indicates index mapping relationship.

3. but noise and abnormal point are inevitable under normal conditions, therefore utilize R₂To estimate x_kAfter rotation One range limit；Make ∠ (Bx_k,y_k)≤ε_i, it is known that Bx_kIt is in Fig. 2 (c) in a hypographous border circular areas of tool.Together Reason, for the point of i ≠ k, can obtain Bx_iAlso in figure (c) in a hypographous border circular areas of tool.

It is denoted as:

2) using the Z axis for the space coordinates built as the second spin matrix R₂Rotary shaft, and R₂=A₂B₂, A₂And B₂Point The second spin matrix R is not indicated₂Split-matrix, B₂By current registration point to collection A_i2And B_i2Middle a pair of registration point is to (x_k,y_k) Meet B₂x_k=y_kWhen, carry out calculating acquisition, A₂Rotary shaft be Z axis, meanwhile, according to the registration distance threshold ξ of setting, calculate It is registrated angle threshold ε_i,

3) rotation byThe border circular areas surrounded, whenThe left margin of the border circular areas surrounded with y_iWhen tangent (shown in such as Fig. 2 (d)), byThe rotation angle of the border circular areas surrounded is θ₁, at this time by The center of circle of the border circular areas surrounded is point M, by point M and y_iDistance r_sA quadratic equation with one unknown is constructed, is existed using point M Equation in great circle O, two equation groups of simultaneous obtain the x and y coordinates value of point M, find out θ₁；

Indicate point x_iBy B₂Rigid body translation is formed with B₂x_iFor the centre of sphere, ε_iFor the border circular areas of radius, such as scheme Shown in 2 (e)；

Because of B₂It is the rigid body translation spin matrix when having noise or abnormal point,It indicates: x_iThrough Cross B₂Rigid body translation, but because there are noise or abnormal point, can only find out a satisfactory range, and this range It is exactly with B₂x_iFor the centre of sphere, ε_iFor the border circular areas of radius；

The great circle O is to give directions x_iIn the space coordinates, using point O as the circle where origin；

4) continue to rotateThe border circular areas surrounded, byThe right margin of the border circular areas surrounded With y_iWhen tangent,The rotation angle of the border circular areas surrounded is θ₂, at this time byThe circle surrounded The center of circle in region is point N, by point N and y_iDistance r_sA quadratic equation with one unknown is constructed, using equation of the N in great circle O, Two equation groups of simultaneous obtain the x and y coordinates value of point N, find out θ₂, obtain angle difference Δ θ_i；

Wherein, r_sIt isThe radius of the border circular areas surrounded, г indicate the B in space coordinates O₂x_iTo point O Distance, r_s=г * tan ε_i。

Specific embodiment described herein is only an example for the spirit of the invention.The neck of technology belonging to the present invention The technical staff in domain can make various modifications or additions to the described embodiments or replace by a similar method In generation, however, it does not deviate from the spirit of the invention or beyond the scope of the appended claims.

Claims (7)

1. a kind of three-dimensional image registration method for screening out error hiding based on multiple dimensioned description, which is characterized in that including following step It is rapid:

Step 1: obtaining the three-dimensional point cloud X and Y of the adjacent scene in visual angle, and multiple dimensioned for each point construction in three-dimensional point cloud X and Y Son (Q is described_i, N_i)；

N_i={ n_l}；

Wherein, 1 indicate that the neighborhood search radius at the midpoint three-dimensional point cloud X or Y marks, value is followed successively by 1,2 ..., and L, L indicate neighborhood Search radius sum, L >=2；T_lIndicate normalized vector,λ_l1, λ_l2, λ_l3For Matrix C_lCarry out SVD points Characteristic value after solution, λ_l1≥λ_l2≥λ_l3, n_l1, n_l2, n_l3Respectively λ_l1, λ_l2, λ_l3Corresponding feature vector,S_l={ p_i|||p_i-p₀||≤r_l, | S_l| indicate S_lMaximum value, p_iIndicate three Certain point p in dimension point cloud X or Y₀Neighborhood point, r_lIndicate point p₀The 1st neighborhood search radius,r_LFor setting Maximum neighborhood search radius, n_lRespectively indicate the point normal vector corresponding with the search radius of neighbourhood 1 of point in three-dimensional point cloud, value For n_l3；

Step 2: multiple dimensioned description of each point in traversal three-dimensional point cloud X and Y will have identical multiple dimensioned description in X and Y Point be recorded as crucial point set A respectively_i0And B_i0；

Step 3: according to principal curvatures signature search key point set A_i0And B_i0In initial matching point set A_i1And B_i1；

From crucial point set A_i0Choose arbitrary point a_i, traverse B_i0Midpoint is selected and point a_iThe point b to match_j, to meet the following conditions Point construct initial matching point set A_i1And B_i1:

Wherein, k₁(a_i) and k₂(a_i) it is illustrated respectively in point a_iPlace, utilizes point a_iNeighborhood point obtain principal curvatures minimum and master Curvature maximum, k₁(b_j) and k₂(b_j) it is illustrated respectively in point b_jPlace, utilizes point b_jNeighborhood point obtain principal curvatures minimum and Principal curvatures maximum, δ₁For the first error threshold value of setting；

Step 4: using similar triangles threshold value to initial matching point set A_i1And B_i1In matching double points screened, be registrated Point is to collection A_i2And B_i2；

Step 5: A is calculated based on SVD_i2And B_i2Between the first spin matrix R₁With the first translation matrix T₁, utilize the first rotation Torque battle array R₁With the first translation matrix T₁, successively calculate A_i2In each point a_i2Change point a_3iCoordinate, while calculating A_i2In it is every A point is in B_i2In match point b_2iWith the Euclidean distance d between correspondent transform point_3i；

A_i3=R₁*A_i2+T₁

Wherein, (a_3ix, a_3iy, a_3iz) and (b_i2x, b_i2y, b_i2z) it is respectively a_3iAnd b_i2Coordinate on x, y, z axis；

Step 6: calculatingThe initial value of k is that 1, N is current registration point to collection A_i2And B_i2Middle registration point pair Number, and judge d_k+1-d_k< ξ₃It is whether true, ξ₃Indicate the error threshold greater than zero, if so, then with current registration point to collection A_i2 And B_i2As the registration data of three-dimensional point cloud X and Y, otherwise, k=k+1 being enabled, return step 3 chooses each neighborhood of a point point again, Calculate principal curvatures.

2. the method according to claim 1, wherein utilizing rotation angle threshold θ_mRegistration is screened again to doing, By current registration point to collection A_i2And B_i2In, the corresponding registration point of following formula is unsatisfactory for as abnormal point or noise spot progress It rejects:

Δθ_i< θ_m

Wherein, Δ θ_iIndicate current registration point to collection A_i2And B_i2Rotation angle difference between middle i-th pair registration point pair；

Current registration point is to collection A_i2And B_i2Middle i-th pair registration point is to (x_i, y_i) between rotation angle difference solution procedure such as Under:

1) from current registration point to collection A_i2And B_i2In, two pairs of non-coplanar registrations pair are randomly selected, with four non-coplanar point structures At a spherical surface, affiliated spherical equation and the centre of sphere 0 are found out by analytic geometry and establish space coordinates with point 0 for origin；

2) using the Z axis for the space coordinates built as the second spin matrix R₂Rotary shaft, and R₂=A₂B₂, A₂And B₂Table respectively Show the second spin matrix R₂Split-matrix, B₂By current registration point to collection A_i2And B_i2Middle a pair of registration point is to (x_k, y_k) full Sufficient B₂x_k=y_kWhen, carry out calculating acquisition, A₂Rotary shaft be Z axis, meanwhile, according to the registration distance threshold ξ of setting, calculating is matched Quasi- angle threshold ε_i,

3) rotation byThe border circular areas surrounded, whenThe left margin and y of the border circular areas surrounded_iPhase When cutting, byThe rotation angle of the border circular areas surrounded is θ₁, at this time byThe border circular areas surrounded The center of circle is point M, by point M and y_iDistance r_sA quadratic equation with one unknown is constructed, equation of the point M in great circle 0, simultaneous are utilized Two equation groups obtain the x and y coordinates value of point M, find out θ₁；

Indicate point x_iBy B₂Rigid body translation is formed with B₂x_iFor the centre of sphere, ε_iFor the border circular areas of radius；

The great circle 0 is to give directions x_iIn the space coordinates, to put 0 circle where origin；

4) continue to rotateThe border circular areas surrounded, byThe right margin and y of the border circular areas surrounded_i When tangent,The rotation angle of the border circular areas surrounded is θ₂, at this time byThe border circular areas surrounded The center of circle is point N, by point N and y_iDistance r_sA quadratic equation with one unknown is constructed, equation of the N in great circle 0, simultaneous two are utilized A equation group obtains the x and y coordinates value of point N, finds out θ₂, obtain angle difference Δ θ_i；

Wherein, r_sIt isThe radius of the border circular areas surrounded, r indicate the B in space coordinates 0₂x_iTo point 0 away from From r_s=r*tan ε_i。

3. the method according to claim 1, wherein in the three-dimensional point cloud each point principal curvatures calculating process It is as follows:

Step 3.1: in three-dimensional point cloud arbitrary point and the neighborhood of a point point be fitted to a local secondary curved surface G (x, y)；

G (x, y)=ax²+bxy+cy²+dx+ey

Step 3.2: derivation being carried out to local quadratic surface G (x, y), obtains first fundamental constant E, F, G and the second fundamental constant L,M,N；

E=G_xx, F=G_x*G_y, G=G_yy；

L=nG_xx, M=nG_xy, N=nG_yy；

Wherein, G_x、G_yRespectively G (x, y) asks x, y and once leads, G_xx、G_yyIt asks secondary to x, y for G (x, y) to lead, G_xyFor G (x, y) Successively to x and y derivation, n is normal vector of the local secondary curved surface G (x, y) at selected arbitrary point；

Step 3.3: calculate Gaussian curvature K and mean curvature H:

Step 3.4: calculating principal curvatures minimum k₁With principal curvatures maximum k₂:

4. the method according to claim 1, wherein described utilize similar triangles threshold value to initial matching point set A_i1And B_i1In the process screened of matching double points it is as follows:

Step 4.1: in A_i1In randomly select 3 not conllinear point a_i, a_j, a_k, in B_i1Find corresponding 3 point b_u, b_v, b_w；

Step 4.2: enabling (a_i, b_u) and (a_j, b_v) it is known matching double points, following calculate is carried out according to Similar Principle of Triangle:

Δa_ij=| | a_i-a_j||；Δa_ik=| | a_i-a_k||；Δa_jk=| | a_j-a_k||；

Δb_uv=| | b_u-b_v||；Δb_uw=| | b_u-b_w||；Δb_wv=| | b_w-b_v||；

Step 4.3: by Similar Principle of Triangle, calculating two triangle similarity Ω；

Wherein, Δ a_ij, Δ a_ik, Δ a_jk；Δb_uv, Δ b_uw, Δ b_wvFor each side length of triangle, δ₂For the second error of setting Threshold value；

Step 4.4: if a_i, a_j, a_k, b_u, b_v, b_wMeet Ω < δ₂, then it is assumed that a_kAnd b_wBelong to a pair of of registration point pair.

5. method according to claim 1-4, which is characterized in that the L value is 4, and r₄=0.004m.

6. according to the method described in claim 4, it is characterized in that, the δ₂Value is any one in 0.02,0.05,0.1 It is a.

7. the method according to claim 1, wherein obtaining the adjacent scene in visual angle using 3D profile scanning sensor Three-dimensional point cloud.