CN100538265C - Quick scene matching aided navigation method based on convolution - Google Patents

Abstract

The invention belongs to technical field of image matching in the Terrain-aided Navigation system, it is characterized in that: in real-time figure and reference diagram matching process, utilize the reference diagram matrix to carry out convolution algorithm and obtain convolution matrix with the backward matrix that obtains by real-time figure matrix, and this convolution matrix is obtained new matrix divided by the summation that real-time figure matrix correspondence position multiplies each other in twos, get this matrix certain limit wherein as the coupling possible position, the position of choosing with numerical value 1 immediate value is final matched position, and this position is the matched position of real-time figure at reference diagram.This method has very significant raising than the arithmetic speed of previous methods.

Description

Quick scene matching aided navigation method based on convolution

Technical field

The invention belongs to technical field of image matching in the Terrain-aided Navigation system.

Background technology

Scene matching aided navigation is taken two width of cloth images that (admission) get off and is spatially aimed at from comprising same scenery (zone) exactly, to determine the process of the relative position relation between this two width of cloth image.It has crucial meaning in fields such as navigational guidance, medical science and data fusion.In its application process, some situation is less demanding to speed, some situation is very high to the requirement of speed, as: in utilizing the cruise missile flight course of Terrain-aided Navigation, even the unhappy subsonic speed of flying speed, if be 0.8 overtone speed, flight 1km also only needed for 3.6765 seconds, may need repeatedly imaging coupling in this short time, therefore the correction of data and the execution of control system etc. need matching algorithm to have very fast arithmetic speed.Normalized crosscorrelation technology (Normalized Cross Correlation Method is called for short NCC) is that a kind of noise resisting ability comparatively speaking is strong, coupling accurate match algorithm, therefore is widely used in navigational guidance.But because its operand is bigger, it is slow comparatively speaking to adopt conventional compute mode to play computing velocity, requires than higher Terrain-aided Navigation system for real-time, and its arithmetic speed needs further to improve.

Summary of the invention

The objective of the invention is to improve the arithmetic speed of normalized crosscorrelation matching algorithm (NCC),, or gain time for the rapid remote sensing data fusion so that carry out Terrain-aided Navigation more quickly.

The invention is characterized in that this method is to realize according to the following steps successively on the computing machine in the Terrain-aided Navigation system:

Step (1) as with reference to figure, is represented the gray level image in a certain area of taking in advance with A, this is imported computing machine with reference to figure A, is m with reference to the size of figure A _A* n _A, the value in the matrix is a gray-scale value; The gray level image in the described area of vehicle upper sensor shooting as scheming in real time, is represented with B this is schemed B in real time import computing machine, the size of scheming B in real time is m _B* n _B, the value in the matrix is a gray-scale value; The span of described gray-scale value is 0～255; And to make the initial position of real-time figure subgraph on reference diagram be x _S=1, y _S=1;

Step (2) is all put upside down the sequencing of the row and column of the matrix of scheming B in real time mutually successively by following formula, obtains Matrix C:

C(m _B-i+1，n _B-j+1)＝B(i，j)，i＝1，2，…，m _B，j＝1，2，…，n _B；

Calculate sum (B.*B) simultaneously, all same position elements multiply each other in twos in symbol " .* " expression two matrixes, and summation is asked in " sum " expression;

Step (3) is made convolution algorithm to matrix and Matrix C with reference to figure A, obtains matrix D, and size is (m _A+ m _B-1) * (n _A+ n _B-1);

Step (4) is got (m in the described matrix D _B, n _B) to (m _A, n _A) between each value as possible matched position, and use matrix D _AExpression;

Step (5) is D _AIn each divided by sum (B.*B), obtain D _AB

Step (6) is found out D _ABIn with the position (x of numerical value 1 immediate value _SB, y _SB), just scheme B in real time with reference to the position among the figure A, and output parameter (x _SB, y _SB).

Advantage of the present invention: the matching speed of the method for the invention improves a lot than existing methods.

Description of drawings

Fig. 1 is the process flow diagram of the inventive method;

Fig. 2 is the coordinate definition mode of image, x _S, y _SIndependent variable for expression subgraph reference position;

Fig. 3 is the invert method and the flow process of preface matrix;

Fig. 4 is method and a flow process of asking the matrix convolution;

Fig. 5 for reference diagram in the example 1 and utilize this method calculate match point and figure and the original match point selected in real time;

Fig. 6 for reference diagram in the example 2 and utilize this method calculate match point and figure and the original match point selected in real time.

Embodiment

1, hardware environment: 586 computing machines (CPU 2.7GHZ), operating system windows xp is equipped with Matlab7.0 software

2, step

1) software arrangements

Title: fas_NCC.m

The position: this software is arranged in navigational guidance system scene matching aided navigation position

Function: improve the arithmetic speed of normalized crosscorrelation matching algorithm, vehicle (aircraft or the guided missile etc.) position of image that upper sensor becomes (scheming B in real time) in reference diagram (A) is provided fast.

Definition: A, reference diagram, the gray level image in a certain area of Pai Sheing in advance

B schemes in real time, the gray level image that the vehicle upper sensor is taken

Step: face as follows

Interface: be input as with reference to figure A (visible images) and real-time figure B (visible light or infrared image), be output as real-time figure B with reference to the coordinate position (x among the figure A _SB, y _SB)

2) coupling step

A. hardware initialization

Parameter is provided with: the gray scale of jpg form is with reference to figure A (big or small m _A* n _A) and gray scale scheme B (big or small m in real time _B* n _B), shape is rectangular, and reference diagram is identical with the resolution of real-time figure, the direction unanimity.

Numerical range: reference diagram (200～10000) * (200～10000)

Scheme (20～1000) * (20～1000) in real time

B. performing step

1. call in (matrix form, big or small m with reference to figure A _A* n _A) and scheme B (matrix form, big or small m in real time _B* n _B), the value in the matrix is the gray level image in somewhere.

Annotate: gray scale is the numerical value of expression brightness size in the image, magnitude range 0～255.

2. real-time figure B is carried out backward computing (just the sequencing of the row and column of matrix is all put upside down mutually) according to the method for Fig. 3, obtain Matrix C, calculate Sum (B.*B) (.* represents that all same position elements multiply each other in twos in two matrixes, and Sum represents all elements sum) simultaneously.

3. according to Fig. 4 A and C are carried out convolution algorithm and obtain matrix D, size is (m _A+ m _B-1) * (n _A+ n _B-1).

4. get (m in the matrix D _B, n _B) to (m _A, n _A) each position be possible matched position, and be designated as D _A

5. D _AIn every obtain D divided by sum (B.*B) _AB

6. find out D _ABIn with the position (x of numerical value 1 immediate value _SB, y _SB) be the matched position of image B, just scheme the position of B in reference diagram in real time, and output parameter (x _SB, y _SB).

Annotate: it is directions X to the right that the coordinate in the image is defined as, and is downwards the Y direction.Fig. 2 (X is corresponding with n, and Y is corresponding with m) is seen in concrete definition.

The ultimate principle of this method is that reference diagram in the images match and real-time figure are changed into two signals, utilize the convolution of signal to realize normalized crosscorrelation images match (NCC), and then utilize rapidity, the globality of matrix convolution, improve the arithmetic speed of NCC greatly.

In order to verify rapidity and the validity of utilizing convolution algorithm, carried out a large amount of l-G simulation tests, and with conventional NCC method (Original Method, in layer calculate according to normal thinking) and optimize after NCC method (FasterMethod, adopt some little skills, utilize .* to carry out the matrix corresponding point and multiply each other) contrast.In the process of the test, the entire image of reference diagram for showing, and the real-time figure of experiment usefulness is the little image that intercepts according to certain window in whole realtime graphic, helps multiple authentication like this.Simulated environment is: 586 computing machines (CPU 2.7GHZ), operating system windows xp is equipped with Matlab7.0 software.

Example 1: reference diagram is a visible images, and size is 200 * 200, and figure is the visible images that has added salt-pepper noise in real time, and the match window size is respectively 30 * 30 and 50 * 50.Fig. 5 for reference diagram and utilize this method calculate match point and in real time figure and the original match point selected.

The experimental result of experiment 1 sees Table 1.

The match time of table 1 example 1

Tab.1?The?matching?time?of?example?1

The unit of match time is second

Example 2: reference diagram is middle infrared image, and size is 500 * 500, and figure is the middle infrared image that has added white noise in real time, and the match window size is respectively 30 * 30 and 50 * 50.Fig. 6 for reference diagram and utilize this method calculate match point and in real time figure and the original match point selected.

The experimental result of example 2 sees Table 2.

The match time of table 2 example 2

Tab.2?The?matching?time?of?example?2

The unit of match time is second

Find following phenomenon by above example:

(1) in matching process, conventional NCC algorithm is maximum operation time, and the NCC algorithm time of optimization is more, and the used time of convolution is minimum, and compares less a lot with other two kinds of methods;

(2) under window 30 * 30 situations, the image to 200 * 200, operation time, ratio was 1:37:48, the image to 500 * 500, operation time, ratio was 1:53:66.To 200 * 200 image, operation time, ratio was 1:27:30 under window 50 * 50 situations, the image to 500 * 500, and operation time, ratio was 1:42:48.Compare with other two kinds of methods as can be seen, big more at image, just under the situation that calculated amount is big more, utilize the convolution technique advantage obvious more, operation time is well below an order of magnitude.

Examples of implementation are as follows:

In order to be illustrated more clearly in problem, illustrate for the example of a simple two-dimensional convolution, establish

$A=\left(\begin{array}{ccccc}1& 2& 3& 7& 43\\ 23& 32& 56& 89& 121\\ 221& 231& 21& 13& 34\\ 56& 112& 71& 90& 89\end{array}\right)$ $B=\left(\begin{array}{cc}56& 89\\ 21& 13\end{array}\right)$ m _A＝4，n _A＝5，m _B＝2，n _B＝2

1. B being carried out the backward computing obtains $C=\left(\begin{array}{cc}13& 21\\ 89& 56\end{array}\right)$ , calculate sum (B.*B)=13*13+89*89+21*21+56*56=11667 (all same position elements multiply each other in twos in the * representing matrix, and Sum represents all elements sum) simultaneously.

2. according to Fig. 4 A and C are carried out convolution algorithm and obtain matrix

$D=\left(\begin{array}{cccccc}13& 47& 81& 154& 706& 903\\ 388& 1133& 1779& 3124& 7661& 4949\\ 4920& 11780& 11900& 11667& 16468& 7490\\ 20397& 35567& 18080& 4994& 6801& 3773\\ 4984& 13104& 12591& 11986& 12961& 4984\end{array}\right).$

3. get (m in the matrix D _B, n _B)=(2,2) to (m _A, n _A)=(4,5)=(4,5) value $\left(\begin{array}{cccc}1133& 1779& 3124& 7661\\ 11780& 11900& 11667& 16468\\ 35567& 18080& 4994& 6801\end{array}\right)$ Be possible matched position, and be designated as D _A

4. D _AIn every obtain divided by Sum (B.*B)

5. find out D _ABIn with numerical value 1 immediate position (x _SB, y _SB)=(3,2) be the matched position of image B, just scheme the position of B in reference diagram (noticing that the position with the B coupling is (3,2)) in real time, and output parameter (x _SB=3, y _SB=2).

Annotate: it is directions X to the right that the coordinate in the image is defined as, and is downwards the Y direction.Fig. 2 (X is corresponding with n, and Y is corresponding with m) is seen in concrete definition.

Claims (1)

1. Based on the rapid image matching method of convolution, it is characterized in that 1, this method is to realize according to the following steps successively on the computing machine in the Terrain-aided Navigation system:

   Step (1) as with reference to figure, is represented the gray level image in a certain area of taking in advance with A, this is imported computing machine with reference to figure A, is m with reference to the size of figure A _A* n _A, the value in the matrix is a gray-scale value; The gray level image in the described area of vehicle upper sensor shooting as scheming in real time, is represented with B this is schemed B in real time import computing machine, the size of scheming B in real time is m _B* n _B, the value in the matrix is a gray-scale value; The span of described gray-scale value is 0～255; And to make the initial position of real-time figure subgraph on reference diagram be x _S=1, y _S=1;

   Step (2) is all put upside down the sequencing of the row and column of the matrix of scheming B in real time mutually successively by following formula, obtains Matrix C:

   C(m _B-i+1，n _B-j+1)＝B(i，j)，i＝1，2，…，m _B，j＝1，2，…，n _B；

   Calculate sum (B.*B) simultaneously, all same position elements multiply each other in twos in symbol " .* " expression two matrixes, and summation is asked in " sum " expression;

   Step (3) is made convolution algorithm to matrix and Matrix C with reference to figure A, obtains matrix D, and size is (m _A+ m _B-1) * (n _A+ n _B-1);

   Step (4) is got (m in the described matrix D _B, n _B) to (m _A, n _A) between each value as possible matched position, and use matrix D _AExpression;

   Step (5) is D _AIn each divided by sum (B.*B), obtain D _AB

   Step (6) is found out D _ABIn with the position (x of numerical value 1 immediate value _SB, y _SB), just scheme B in real time with reference to the position among the figure A, and output parameter (x _SB, y _SB).

Images