CN103021021B - Adopt the generalized stereopair three-dimensional rebuilding method of variance components estimate - Google Patents

Abstract

The generalized stereo pair 3D reconstruction method using variance component estimation involves a 3D reconstruction method, in order to solve the resolution difference of each pixel of the same image and the resolution difference between images under different imaging conditions due to large angles under different imaging systems. However, the existing RFM-based generalized stereo pair 3D reconstruction methods have the problem of severely unbalanced generalized linear equations. The method is to realize the method of embedding the variance component estimation method into the classical three-dimensional reconstruction method based on the forward intersection of the rational polynomial model RFM by using the pixel-level resolution relationship of the corresponding target point to be reconstructed in the generalized stereo image pair to realize iterative solution to the unknown The weight matrix in the process of three-dimensional coordinates is adaptively adjusted, thereby effectively improving the accuracy of three-dimensional reconstruction and improving the feasibility of three-dimensional reconstruction of generalized stereo pairs, which is used for three-dimensional reconstruction of generalized stereo pairs.

Description

Generalized stereopair three-dimensional reconstruction method adopting variance component estimation

Technical Field

The invention relates to a three-dimensional reconstruction method, in particular to a generalized stereopair three-dimensional reconstruction method adopting variance component estimation.

Background

With the rapid development of remote sensing technology, sensor technology and computer technology, human beings realize unprecedented striding development in the field of ground observation. More and more observation platforms such as high-resolution remote sensing commercial satellites, military satellites, near space aircrafts, aviation airplanes and unmanned reconnaissance planes enable people to obtain large-range high-precision three-dimensional scene models.

The remote sensing data is utilized to carry out three-dimensional reconstruction, and the research hotspot in the field of remote sensing is formed. In the early stage, people utilize a strict sensor model, accurately acquire the internal and external orientation parameters of an imaging sensor, and through the front intersection of multi-angle stereopair acquired by the same imaging system, the three-dimensional reconstruction of an interested target point is realized by utilizing a least square technology, and further, the rapid generation of a large-range scene DEM can be realized by utilizing an interpolation technology; with the provision of Rational Function Model (RFM) of satellite images for users by high resolution satellite service providers such as IKONOS2, QuickBird, etc., three-dimensional reconstruction techniques based on RFM have been extensively studied. In general, the development of three-dimensional reconstruction techniques based on RFM has gone through three stages: the first stage is to construct a least square estimation model by utilizing a forward-convergence stereo imaging technology and solve unknown three-dimensional coordinates by adopting given RFM parameters; in the second stage, it is realized that the original RFM model may have a relatively serious systematic error, and for this reason, the three-dimensional reconstruction accuracy can be improved by two approaches. The first method is to realize the correction of the original RFM model by utilizing a Ground Control Point (GCP) with accurate three-dimensional coordinates and adopting Bundle-Adjustment technology and the like; the second method does not require modification of the RFM model, but also achieves improvement of the accuracy of three-dimensional reconstruction of the target points by calculating a coordinate transformation model (modifying the input or output in the original RFM forward merge method) in the image domain or the target domain using the GCP with accurate three-dimensional coordinates. In the third stage, in order to better utilize wide remote sensing data, people have not been constrained to acquire stereo data under the same imaging system, but have realized three-dimensional reconstruction by jointly utilizing remote sensing data with different resolutions of different imaging systems according to requirements. But the currently used methods also only stay on a small complement to the classical RFM-based forward-convergence three-dimensional reconstruction method.

However, whether the three-dimensional reconstruction method based on a strict sensor model or the three-dimensional reconstruction method based on the RFM, most of the existing methods strictly require that the imaging sensors forming the stereopair images have strictly identical internal and external orientation parameters except for the imaging pitch angle, so that the stereopair has completely identical dimensional relationship with respect to the overlapped region. This results in that the remote sensing data under different imaging conditions for the same observation scene, such as different platforms, different angles, and different time phases, cannot be fully utilized. Therefore, there is a need to study the relevant theory to realize the three-dimensional reconstruction task by using the generalized stereopair composed of the remote sensing images with the overlapped areas obtained by different imaging systems. In the face of the situation that resolution difference of pixels in the same image and resolution difference between images under different imaging conditions are often very large due to large angles in different imaging systems, a classical mathematical model based on an RFM (recursive RFM) forward intersection three-dimensional reconstruction method becomes a generalized linear equation set with serious imbalance. Therefore, there is an urgent need to fundamentally solve the problem of three-dimensional reconstruction of a generalized stereopair with severe fractional imbalance.

Disclosure of Invention

The invention aims to solve the problem that an existing RFM-based generalized stereopair three-dimensional reconstruction method has a generalized linear equation set with serious imbalance when the resolution difference of each pixel of the same image and the resolution difference between images under different imaging conditions are large due to large angles under different imaging systems.

The invention relates to a generalized stereopair three-dimensional reconstruction method adopting variance component estimation, which comprises the following steps:

the method comprises the following steps: according to the parameters of the imaging sensor, obtaining the pixel level resolution of a corresponding point of a target point to be reconstructed in each image of the generalized stereopair;

step two: constructing a generalized linear equation set related to the three-dimensional reconstruction of a target point to be reconstructed by utilizing the RFM universal sensor model of each image of the generalized stereopair;

step three: initializing a weight matrix of the generalized linear equation set constructed in the second step by using the pixel level resolution obtained in the first step;

step four: performing initial estimation on the three-dimensional coordinates of a target point to be reconstructed by using weighted least square estimation;

step five: according to the current weight matrix and the initial value of the three-dimensional coordinate of the target point to be reconstructed, obtaining the increment of the three-dimensional coordinate of the target point to be reconstructed under the current iteration relative to the initial value of the three-dimensional coordinate by utilizing weighted least square estimation, and obtaining the currently estimated error amount;

step six: judging whether the currently estimated error quantity meets the precision requirement, if so, outputting an estimated value of the three-dimensional coordinates of the current target point to be reconstructed to complete three-dimensional reconstruction, and if not, turning to the seventh step;

step seven: and updating the current weight matrix by using a variance component estimation method according to the current estimated error amount, updating the estimated value of the three-dimensional coordinate of the target point to be reconstructed by using weighted least square estimation, and turning to the fifth step.

The method has the advantages that the problem of serious unbalance of a RFM-based generalized stereopair three-dimensional reconstruction mathematical model is solved by analyzing the actual resolution of the target point pixel in each image of the generalized stereopair in detail and combining variance component estimation, and the precision of realizing the three-dimensional reconstruction of the generalized stereopair by a classical RFM-based three-dimensional reconstruction algorithm is effectively improved. The generalized stereopair with different imaging conditions can be comprehensively utilized to realize the three-dimensional reconstruction task of the interested target point, and the utilization efficiency of the existing data is improved.

Drawings

Fig. 1 is a schematic flow chart of a generalized stereopair three-dimensional reconstruction method using variance component estimation according to the present invention.

Fig. 2 is a schematic diagram illustrating the analysis of pixel-level resolution of a generalized stereopair three-dimensional reconstruction method using variance component estimation according to a second embodiment of the present invention, where a denotes an imaging sensor.

Fig. 3 is a schematic diagram illustrating analysis of pixel-level resolution of a generalized stereopair three-dimensional reconstruction method using variance component estimation according to a third embodiment of the present invention, where a denotes an imaging sensor.

Fig. 4 is a schematic diagram of control points of a multi-temporal high-resolution satellite image acquired by a QuickBird satellite, which uses the generalized stereopair three-dimensional reconstruction method using variance component estimation according to the present invention.

Fig. 5 is a schematic diagram of control points of a multi-temporal high-resolution satellite image acquired by a WorldView2 satellite, which is obtained by the generalized stereopair three-dimensional reconstruction method using variance component estimation according to the present invention.

Detailed Description

The first embodiment is as follows: the embodiment is described with reference to fig. 1, the generalized stereopair three-dimensional reconstruction method using variance component estimation according to the embodiment,

it comprises the following steps:

the method comprises the following steps: according to the parameters of the imaging sensor, obtaining the pixel level resolution of a corresponding point of a target point to be reconstructed in each image of the generalized stereopair;

step two: constructing a generalized linear equation set related to the three-dimensional reconstruction of a target point to be reconstructed by utilizing the RFM universal sensor model of each image of the generalized stereopair;

step three: initializing a weight matrix of the generalized linear equation set constructed in the second step by using the pixel level resolution obtained in the first step;

step four: performing initial estimation on the three-dimensional coordinates of a target point to be reconstructed by using weighted least square estimation;

step five: according to the current weight matrix and the initial value of the three-dimensional coordinate of the target point to be reconstructed, obtaining the increment of the three-dimensional coordinate of the target point to be reconstructed under the current iteration relative to the initial value of the three-dimensional coordinate by utilizing weighted least square estimation, and obtaining the currently estimated error amount;

step six: judging whether the currently estimated error quantity meets the precision requirement, if so, outputting an estimated value of the three-dimensional coordinates of the current target point to be reconstructed to complete three-dimensional reconstruction, and if not, turning to the seventh step;

step seven: and updating the current weight matrix by using a variance component estimation method according to the current estimated error amount, updating the estimated value of the three-dimensional coordinate of the target point to be reconstructed by using weighted least square estimation, and turning to the fifth step.

The initialized weight matrix in the third step is a weight matrix required in the process of carrying out weighted least square estimation solving; the method mainly aims to eliminate the problem of imbalance caused by different image resolutions according to the resolution relationship between stereopair images and provide a better initial value for estimating the optimal variance component in the subsequent iteration

The second embodiment is as follows: this embodiment is a further limitation on the method for three-dimensional reconstruction of a generalized stereopair using variance component estimation as described in the first embodiment,

in the first step, according to the parameters of the imaging sensor, the method for obtaining the pixel level resolution of the corresponding point of the target point to be reconstructed in each image of the generalized stereopair comprises the following steps:

according to the height H, the instantaneous field angle theta, the pitch angle beta and the image size N × N of the imaging sensor, the resolution of the kth pixel in the image obtained by the sensor is obtained as follows:

will be the resolution Res_kAnd the resolution of the target point to be reconstructed in the pixel level of the corresponding point in each image of the generalized stereopair, wherein N is a positive integer.

The present embodiment is based on the actual situation of acquiring data, and firstly, if the imaging parameters of the imaging sensor can be acquired in great detail, as shown in fig. 2. In the figure, N represents the total number of pixels. The distance from the front view subsatellite point o to the point k in fig. 2 is:

$\begin{array}{c}L=H\·tan(k\frac{2\mathrm{\β}}{N}+\mathrm{\θ}-\mathrm{\β})\\ =H\·tan(\mathrm{\θ}+\frac{2k-N}{N}\mathrm{\β})\end{array}---\left(1\right)$

wherein,thus, the resolution of the kth pixel is the distance from the front view intersatellite point o to point k in FIG. 2 minus the distance to point k-1):

$\begin{array}{c}{\mathrm{Res}}_k=H\·\tan (\mathrm{\θ}+\frac{2k-N}{N}\mathrm{\β})-H\·\tan (\mathrm{\θ}+\frac{2k-N-2}{N}\mathrm{\β})\\ =H\·\left(\tan \right(\mathrm{\θ}+\frac{2k-N}{N}\mathrm{\β})-\tan (\mathrm{\θ}+\frac{2k-N-2}{N}\mathrm{\β}\left)\right)\end{array}---\left(2\right)$

the above is performed when the angle θ is greater than 0, and when the angle θ is smaller than zero, the above formula (2) is simply modified as follows:

${\mathrm{Res}}_k=H\·\left(tan\right(\left|\mathrm{\θ}+\frac{2k-N-2}{N}\mathrm{\β}\right|)-tan(\left|\mathrm{\θ}+\frac{2k-N}{N}\mathrm{\β}\right|\left)\right)---\left(3\right)$

the combined formula (2) and formula (3) have:

the third concrete implementation mode: this embodiment is a further limitation on the method for three-dimensional reconstruction of a generalized stereopair using variance component estimation as described in the first embodiment,

in the first step, according to the parameters of the imaging sensor, the method for obtaining the pixel level resolution of the corresponding point of the target point to be reconstructed in each image of the generalized stereopair comprises the following steps: average resolution Res of an orthographic image according to a known imaging sensor_nand a pitch angle beta, and obtaining the pixel level resolution of the corresponding point of the target point to be reconstructed in each image of the generalized stereopair as ${\mathrm{Res}}_{\mathrm{\β}}=\frac{{\mathrm{Res}}_n}{{\cos }^2\mathrm{\β}}.$

When only the condition of the imaging sensor average resolution and its imaging pitch angle is known, as shown in fig. 3;

as can be seen from FIG. 3, the average resolution of the front-view image is Res_naverage resolution Res of image at imaging angle beta_nAnd Res_βThe following geometrical relationship exists between the two components:

${\mathrm{Res}}_{\mathrm{\β}}=\frac{{\mathrm{Res}}_n}{{\cos }^2\mathrm{\β}}---\left(5\right)$

the fourth concrete implementation mode: this embodiment is a further limitation on the method for three-dimensional reconstruction of a generalized stereopair using variance component estimation as described in the first embodiment,

in the second step, the generalized linear equation set for the three-dimensional reconstruction of the target point to be reconstructed is as follows:

$\stackrel{\·\·}{Y}=\stackrel{\·\·}{A}\mathrm{\γ}+\stackrel{\·\·}{E}\⇔\left\{\begin{array}{c}{Y}_1=A_1\mathrm{\γ}+E_1\\ Y_2=A_2\mathrm{\γ}+E_2\\ .\\ .\\ .\\ Y_l=A_l\mathrm{\γ}+E_l\end{array}\right.,$ wherein l is the number of the platforms for obtaining the observed value, and the above-mentioned l equations are arranged from low to high according to the resolution of the corresponding platform of the above-mentioned equation, namely equation Y₁Resolution of the corresponding platform is lowest, equation Y_lThe resolution of the corresponding platform is highest; l is greater than 2, and the ratio of the total of the components, $\stackrel{\·\·}{A}=\[\begin{array}{cccc}{A}_1& A_2& ...& A_l\end{array}\]$ to design a matrix;

$\stackrel{\·\·}{E}=\[\begin{array}{cccc}{E}_1& E_2& ...& E_l\end{array}\]$ in order to measure the error of the measurement,

$E_i={\left[\begin{array}{cccc}{e}_1& e_2& ...& e_{n_i}\end{array}\right]}^T,E_i\∈R^{n_i\×1},$ a measurement error corresponding to the ith observation value, and

E_iand E_jNot correlation (i ≠ j1 ≤ i, j ≤ l), E_iAnd E_jFor measuring errors $\stackrel{\·\·}{E}=\[\begin{array}{cccc}{E}_1& E_2& ...& E_l\end{array}\]$ The component (b) of (a) is,

is the i-th platform observation variance, n_iThe number of observations obtained for the ith platform;

$\stackrel{\·\·}{Y}=\[{\begin{array}{cccc}{Y}_1& Y_2& ...& Y\end{array}}_l\]$ is a vector of observations, wherein $Y_i=\[\begin{array}{cccc}{y}_{i1}& y_{i2}& ...& y_{{\mathrm{in}}_i}\end{array}\]$ Is the observed value of the ith platform, i is 1,2, …, l; the covariance matrix of the residual amounts is

The design matrix (design matrix) is a dedicated matrix in weighted least squares estimation, well known to those skilled in the art;

in order to effectively solve the problem of three-dimensional reconstruction of a multisource generalized stereopair based on an RFM sensor model, a variance component estimation technology is embedded into a three-dimensional reconstruction solving model, and then analysis must be carried out from a mathematical model of the root of the model, namely a weighted generalized least square model. For a generalized system of linear equations:

Y＝Aγ+E(6)

wherein Y ═ Y₁y₂…y_n]^T,Y∈R^n×1is a measured value, A ∈ R^n×mDesigning a matrix for a column full rank; gamma-gamma₁γ₂…γ_m]^T,γ∈R^m×1Is an unknown vector to be solved; e ═ E₁e₂…e_n]^T,E∈R^n×1To measure the error.

If all measurements are taken from the same platform, the mathematical expectation of the error vector is E { E } ═ o ═ 00 … 0]^T∈R^n×1(ii) a The covariance matrix of the error vector is Q_E＝Q_Y＝σ²I_n. This case is called equal precision (EQ) mode. Otherwise, if the measured values are obtained from different platforms, it is called an unequal precision (UEQ) mode, and at this time, equation (6) may be expressed in the form of a plurality of EQ modes as follows:

$\left\{\begin{array}{c}{Y}_1=A_1\mathrm{\γ}+E_1\\ Y_2=A_2\mathrm{\γ}+E_2\\ ...\\ Y_l=A_1\mathrm{\γ}+E_l\end{array}\right.---\left(7\right)$

at this time, equation (7) can be expressed as the following standard form:

$\stackrel{\·\·}{Y}=\stackrel{\·\·}{A}\mathrm{\γ}+\stackrel{\·\·}{E}---\left(8\right)$

wherein, $\stackrel{\·\·}{Y}=\[\begin{array}{cccc}{Y}_1& Y_2& ...& Y_l\end{array}\],y=\[\begin{array}{cccc}{y}_{i1}& y_{i2}& ...& y_{{\mathrm{in}}_i}\end{array}\],$ the mathematical expectation of (c) is zero, and its covariance matrix is:

to this end, the generalized least squares estimate of equation (8) is:

$\widetilde{\mathrm{\γ}}={\left({\stackrel{\·\·}{A}}^T{Q}_{\stackrel{\·\·}{E}}^{-1}\stackrel{\·\·}{A}\right)}^{-1}{\stackrel{\·\·}{A}}^T{Q}_{\stackrel{\·\·}{E}}^{-1}\stackrel{\·\·}{Y}---\left(9\right)$

wherein the design matrixAnd measured valueIt is known that, in order to obtain the least squares estimate of equation (8), it is necessary to obtain an accurate completion pair covariance matrixIs estimated.

Is out of order

Wherein,known weight factors or initialized weight factors. Described in matrix form, there are:

$Q=\underset{i=1}{\overset{n}{\Σ}}{U}_i{\mathrm{\σ}}_i^2=\underset{i=1}{\overset{n}{\Σ}}{U}_i{\mathrm{\θ}}_i---\left(11\right)$

wherein, $U_i=\left[\begin{array}{ccccc}0& ...& 0& ...& 0\\ .& & .& & .\\ .& ...& .& ...& .\\ .& & .& & .\\ 0& ...& P_i& ...& 0\\ .& & .& & .\\ .& ...& .& ...& .\\ .& & .& & .\\ 0& ...& 0& ...& 0\end{array}\right]$

then, the above problem of accurately estimating the covariance matrix is transformed into how to accurately obtain the variance component in equation (11)The platform resolution is the average resolution at which the image platform is acquired.

The fifth concrete implementation mode: the present embodiment further defines the generalized stereopair three-dimensional reconstruction method using variance component estimation described in the first embodiment, and the weight matrix initialized in the third step is:

wherein,to be reconstructedThe line resolution ratio of the corresponding pixels in the lowest resolution image and the ith higher resolution image of the generalized stereo image pair,and the column resolution ratio of corresponding pixels of the target point to be reconstructed in the lowest resolution image and the ith higher resolution image in the generalized stereo image pair is obtained.

The sixth specific implementation mode: this embodiment is a further limitation on the method for three-dimensional reconstruction of a generalized stereopair using variance component estimation as described in the first embodiment,

in the seventh step, according to the current error amount, a method for updating the current weight matrix by using a variance component estimation method comprises the following steps:

step seven one: according to the currently estimated error amount $v_k=\stackrel{\·\·}{A}{\hat{X}}_k-\stackrel{\·\·}{Y}={\left[\begin{array}{cccccc}{v}_k^1& v_k^2& ...& v_k^i& ...& v_k^n\end{array}\right]}^T,$ Calculating intermediate variables $r_i=n_i-\frac{tr\left\{\stackrel{\·\·}{A}{\left({\stackrel{\·\·}{A}}^T{Q}_{\stackrel{\·\·}{E}}^{-1}{|}_k\stackrel{\·\·}{A}\right)}^{-1}{\stackrel{\·\·}{A}}^T{P}_i^{-1}\right\}}{{\widehat{\mathrm{\θ}}}_k^i};$

Step seven and two: computing the respective difference components ${\widehat{\mathrm{\θ}}}_{k+1}^i={\widehat{\mathrm{\σ}}}_{k+1}^2=\frac{{\left(v_k^i\right)}^T{P}_i^{-1}{v}_k^i}{r_i};$

Step seven and three: the updated weight matrix:

wherein,representing the amount of error for the k-th iteration,represents the variance component, P, of the ith platform observation in the kth iteration_iAnd expressing a weighted value corresponding to the ith platform observed value in the current weighted matrix.

In the present embodiment, VCE technology is embedded to realize the estimation of the variance component, and the accurate estimation of the unknown quantity γ in the formula (8) is completed. The specific process is as follows:

starting from the first loop (k ═ 1), and stopping the iteration until the precision requirement is met, the following operations are performed in sequence:

$v_k=\stackrel{\·\·}{A}{\hat{X}}_k-\stackrel{\·\·}{Y}={\left[\begin{array}{cccccc}{v}_k^1& v_k^2& ...& v_k^i& ...& v_k^n\end{array}\right]}^T$

$T_k=v_k^T{v}_k$

$r_i=n_i-\frac{tr\left\{\stackrel{\·\·}{A}{\left({\stackrel{\·\·}{A}}^T{Q}_{\stackrel{\·\·}{E}}^{-1}{|}_k\stackrel{\·\·}{A}\right)}^{-1}{\stackrel{\·\·}{A}}^T{P}_i^{-1}\right\}}{{\widehat{\mathrm{\θ}}}_k^i}$

${\widehat{\mathrm{\θ}}}_{k+1}^i=\frac{{\left(v_k^i\right)}^T{P}_i^{-1}{v}_k^i}{r_i}$

${\hat{X}}_{k+1}={\left({\stackrel{\·\·}{A}}^T{Q}_{\stackrel{\·\·}{E}}^{-1}{|}_{k+1}\stackrel{\·\·}{A}\right)}^{-1}{Q}_{\stackrel{\·\·}{E}}^{-1}{|}_{k+1}\stackrel{\·\·}{Y}$

wherein v is_kRepresenting the residual of the kth iteration; t is_kIf the current precision is smaller than the threshold precision, stopping iteration, and otherwise, continuing; r is_iIs an intermediate variable;andrespectively representing the estimated values of the ith variance component in the kth iteration and the kth +1 iteration;the covariance matrix required for the (k + 1) th iteration.

In order to verify the effectiveness of the method of the invention, the invention was verified by constructing a generalized stereopair using real satellite images with different time phases acquired by different imaging systems. Experimental results show that the method has the advantages that the pixel resolution of the remote sensing images under different imaging conditions is fundamentally analyzed, the variance component estimation technology is embedded into the solving process of carrying out three-dimensional reconstruction by utilizing the generalized stereopair, the reconstruction precision of the classic RFM-based three-dimensional reconstruction method is effectively improved, and the utilization rate of remote sensing data under different imaging conditions is improved.

In the following description of the experiments, the method of the invention is denoted AW-RFM, an abbreviation for adaptiveWeightRFM-based method.

The test images used in the experiment are multi-temporal high-resolution satellite images acquired by QuickBird and WorldView2 satellites, corresponding RFM model parameters are known, and 10 ground control points with high-precision elevations (0.06m precision) are selected for algorithm reconstruction effect evaluation, as shown in FIG. 4 and FIG. 5.

In order to verify the performance of the method, the initialization weight factor of the formula (10) in the method is taken as:

$P=\left[\begin{array}{cccc}1& 0& 0& 0\\ 0& 1& 0& 0\\ 0& 0& {\left({\mathrm{\ρ}}_r\right)}^{\mathrm{\α}}& 0\\ 0& 0& 0& {\left({\mathrm{\ρ}}_c\right)}^{\mathrm{\α}}\end{array}\right],\mathrm{\α}\∈\[\begin{array}{cc}1.0& 2.0\end{array}\]\⋐R---\left(12\right)$

where ρ is_rAnd ρ_cThe ratio of the row resolution to the column resolution of the low resolution observation to the high resolution observation. Note: the premise is that the sequence number of the low-resolution observed value is in front of the sequence number of the high-resolution observed value in the formula (7).

In the experiment, the parameter α was chosen to be 2.0, 1.8, 1.6, 1.4, 1.2 and 1.0, respectively, and expressed as AW-RFM20, AW-RFM18, … and AW-RFM10, respectively. The comparison with the classical RFM-based method (C-RFM) was performed using these 6 inventive methods with different parameters.

It should be noted that, because the horizontal coordinate precision of the ground control point is limited, the experiment only gives the result of elevation reconstruction, and in order to reflect the comparison effect, the comparison data of the elevation difference of 10 control points relative to the first control point is given at the same time, and the reconstruction result of each algorithm in the experiment is higher than the elevation of the control point, that is, the elevation difference is in the same direction, and the comparability is very strong.

The results of the experiment are shown in table 1:

table 1: mean square error of elevation reconstruction (unit: meter)

	C-RFM	AW-RFM10	AW-RFM12	AW-RFM14	AW-RFM16	AW-RFM18	AW-RFM20
								Elevation (L)	1.0863	0.9969	1.1396	1.0987	1.0812	1.0653	2.3549
Height difference	1.1595	1.0993	1.1998	1.1687	1.1557	1.1443	2.3094

The above description is only an embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can understand that the modifications or substitutions within the technical scope of the present invention should be covered within the scope of the claims of the present invention.

The invention discloses a problem that the three-dimensional reconstruction of a generalized stereopair cannot be effectively realized by a classic RFM-based three-dimensional reconstruction method and simple deformation thereof, and finds that the critical problem that the three-dimensional reconstruction of the generalized stereopair cannot be effectively realized by the classic RFM-based three-dimensional reconstruction method is not capable of effectively balancing the serious imbalance problem of a three-dimensional reconstruction mathematical model caused by the resolution difference which cannot be ignored between the generalized stereopairs through the research of different development stages and related mathematical models of the conventional classic three-dimensional reconstruction method which utilizes the forward intersection of the remote sensing stereopair and the pixel level resolution analysis of remote sensing images under different imaging conditions. Therefore, the method provided by the invention has the advantages that through the pixel level resolution analysis of the remote sensing images under different conditions, the Variance Component Estimation (VCE) technology is embedded into the iteration process of the classic RFM-based three-dimensional reconstruction solving, the effective estimation of the variances of the observed values with different resolutions in the iteration process is effectively realized, the problem of serious imbalance of mathematical models caused by different resolutions is effectively balanced, the method for effectively utilizing the generalized stereopair to carry out interested target point three-dimensional reconstruction is realized, and the precision of the classic RFM-based three-dimensional reconstruction method is effectively improved.

In consideration of the fact that the classical three-dimensional reconstruction method based on RFM in the remote sensing field mostly uses images with the same or very similar resolution and smaller convergence angle to form a stereopair, in practical application, images with different resolutions and larger convergence angle are often required to form a generalized stereopair to realize the three-dimensional reconstruction of an interested target point so as to solve the problem that a single-platform stereopair cannot be obtained or the three-dimensional reconstruction accuracy of the single-platform stereopair is not enough. The invention provides a generalized stereopair three-dimensional reconstruction method adopting variance component estimation. The method is suitable for realizing three-dimensional reconstruction of any generalized stereopair capable of obtaining the RFM sensor model. Compared with the classic RFM-based three-dimensional reconstruction method, the method provided by the invention can effectively improve the three-dimensional reconstruction precision of the generalized stereopair. Fig. 1 shows a block diagram of the present invention. The key technical contents of the invention comprise two parts of pixel-level image resolution analysis and unbalanced weighted least square estimation solving realized by using a variance component estimation technology.

Claims (5)

1. adopt the generalized stereo pair three-dimensional reconstruction method of variance component estimation, it is characterized in that, it comprises the steps: Step 1: According to the parameters of the imaging sensor, obtain the pixel-level resolution of the corresponding point of the target point to be reconstructed in each image of the generalized stereo pair; Step 2: Utilize the RFM universal sensor model of each image of the generalized stereo image to construct a generalized linear equation system about the three-dimensional reconstruction of the target point to be reconstructed; Step 3: Use the pixel-level resolution obtained in Step 1 to initialize the weight matrix of the generalized linear equations constructed in Step 2; Step 4: Initially estimate the three-dimensional coordinates of the target point to be reconstructed by weighted least square estimation; Step 5: According to the current weight matrix and the initial value of the three-dimensional coordinates of the target point to be reconstructed, use weighted least squares estimation to obtain the increment of the three-dimensional coordinates of the target point to be reconstructed in the current iteration relative to the initial value of the three-dimensional coordinates , and get the current estimated error amount; Step 6: Judging whether the currently estimated error amount meets the accuracy requirement, if so, then output the estimated value of the 3D coordinates of the current target point to be reconstructed, and complete the 3D reconstruction, if not, then go to step 7; Step 7: According to the current estimated error amount, use the variance component estimation method to update the current weight matrix, and then use the weighted least squares estimation to update the estimated value of the three-dimensional coordinates of the target point to be reconstructed, and then go to step Fives; The weight matrix initialized in step 3 is: in, is the line resolution ratio of the target point to be reconstructed in the lowest resolution image and the i-th higher resolution image in the generalized stereo pair, 2≤i≤l, is the column resolution ratio of the target point to be reconstructed in the lowest resolution image and the i-th higher resolution image in the generalized stereo pair. 2. the generalized stereo pair three-dimensional reconstruction method that adopts variance component estimation according to claim 1, it is characterized in that, according to the parameter of imaging sensor in the step 1, obtain the target point to be reconstructed corresponding in each image of generalized stereo pair The method for pixel-level resolution of points is: According to the height H of the imaging sensor, the instantaneous field of view angle θ, the pitch angle β and the image size N*N, the resolution of the kth pixel in the image obtained by the sensor is: The resolution Res _k is taken as the pixel-level resolution of the corresponding point of the target point to be reconstructed in each image of the generalized stereo pair, and N is a positive integer. 3. the generalized stereo image pair three-dimensional reconstruction method that adopts variance component estimation according to claim 1, is characterized in that, In step 1, according to the parameters of the imaging sensor, the method of obtaining the pixel-level resolution of the corresponding point of the target point to be reconstructed in each image of the generalized stereo pair is: according to the average resolution Res _n of the front-view image of the known imaging sensor and Pitch angle β, the pixel-level resolution of the target point to be reconstructed in each image of the generalized stereo pair is obtained as ${\mathrm{Res}}_{\mathrm{\β}}=\frac{{\mathrm{Res}}_{\mathrm{no}}}{{\cos }^2\mathrm{\β}}.$ 4. the generalized stereo image pair three-dimensional reconstruction method that adopts variance component estimation according to claim 1, is characterized in that, the generalized linear equation group of the target point three-dimensional reconstruction to be reconstructed described in step 2 is: $\stackrel{\·\·}{Y}=\stackrel{\·\·}{A}\mathrm{\γ}+\stackrel{\·\·}{\mathrm{E.}}\⇔\left\{\begin{array}{c}{Y}_1=A_1\mathrm{\γ}+{\mathrm{E.}}_1\\ Y_2=A_2\mathrm{\γ}+{\mathrm{E.}}_2\\ \&Center\; Dot;\\ \&Center\; Dot;\\ \&Center\; Dot;\\ Y_l=A_l\mathrm{\γ}+{\mathrm{E.}}_l\end{array}\right.,$ Among them, l is the number of platforms for obtaining observation values, and the above-mentioned l equations are arranged according to the resolution of the platforms corresponding to the equations from low to high, that is, the resolution of the platform corresponding to equation Y ₁ is the lowest, and the resolution of the platform corresponding to equation Y ₁ is the lowest. Highest resolution; l is greater than 2, is the design matrix, γ is the unknown vector to be sought; $\stackrel{\·\·}{\mathrm{E.}}=\left[\begin{array}{cccc}{\mathrm{E.}}_1& {\mathrm{E.}}_2& ...& {\mathrm{E.}}_l\end{array}\right]$ is the measurement error, ${\mathrm{E.}}_i={\left[\begin{array}{cccc}{e}_1& e_2& ...& e_{{\mathrm{no}}_i}\end{array}\right]}^T,{\mathrm{E.}}_i\∈R^{{\mathrm{no}}_i\×1},$ is the measurement error corresponding to the ith observed value, and E _i is not related to E _j , i≠j, 1≤i≤l, 1≤j≤l, the E _i and E _j are measurement errors weight, is the variance of observations on the i-th platform, and n _i is the number of observations obtained on the i-th platform; is the observation vector, where is the observed value of the i-th platform, i=1,2,...,l; the covariance matrix of the measurement error is 5. the generalized stereo pair three-dimensional reconstruction method that adopts variance component estimation according to claim 4, is characterized in that, according to described current error amount in the step 7, utilizes variance component estimation method to update current weight matrix The method is: Step 71: According to the current estimated error amount Calculate intermediate variables $r_i={\mathrm{no}}_i-\frac{tr\left\{\stackrel{\·\·}{A}{\left({\stackrel{\·\·}{A}}^T{Q}_{\stackrel{\·\·}{\mathrm{E.}}}^{-1}{|}_k\stackrel{\·\·}{A}\right)}^{-1}{\stackrel{\·\·}{A}}^T{P}_i^{-1}\right\}}{{\widehat{\mathrm{\θ}}}_k^i},$ The covariance matrix required for the kth iteration; Step 72: Calculate the difference of each party Step 73: The updated weight matrix: in, Indicates the error amount of the kth iteration, Indicates the variance component of the i-th platform observation value in the k-th iteration, and P _i represents the weighted value corresponding to the i-th platform observation value in the current weight matrix.