Multispectral camera dynamic stereo calibration algorithm based on saliency features

Abstract

A multispectral camera dynamic stereo calibration algorithm is based on saliency features. The joint self-calibration method comprises the following steps: step 1: conducting de-distortion and binocular correction on an original image according to internal parameters and original external parameters of an infrared camera and a visible light camera. Step 2: Detecting the saliency of the infrared image and the visible light image respectively based on a histogram contrast method. Step 3: Extracting feature points on the infrared image and the visible light image. Step 4: Matching the feature points extracted in the previous step. Step 5: judging a feature point coverage area. Step 6: correcting the calibration result. The present invention solves the change of a positional relationship between an infrared camera and a visible light camera due to factors such as temperature, humidity and vibration.

Description

TECHNICAL FIELD

The present invention belongs to the field of image processing and computer vision, and relates to a multispectral camera dynamic stereo calibration algorithm based on saliency features.

BACKGROUND

Infrared is an electromagnetic wave with a wavelength between microwave and visible light, and its wavelength is longer than that of red light. Substances higher than absolute zero (−273.15° C.) can generate infrared rays. Infrared images are widely used in different fields such as military and national defense, resource exploration, weather forecasting, environmental monitoring, medical diagnosis and treatment and marine research due to the capability of observation through fog and rain. The infrared can be used to shoot scenes through mist and smoke, and can also be used for infrared photography at night. An infrared camera has the advantage of imaging in extreme scenes (low light, rain, snow and dense fog), and has the disadvantages of low resolution and blurred image details. In contrast, a visible light camera has the advantages of high resolution and clear image details, but cannot be used for imaging in the extreme scenes. Therefore, the combination of the infrared camera and the visible light camera has great practical significance.

Stereo vision is an important topic in the field of computer vision. The purpose is to reconstruct the 3D geometric information of the scenes. Binocular stereo vision is an important field of stereo vision. In the binocular stereo vision, left and right camera lenses are used to simulate two eyes. Depth images are calculated by calculating the difference between binocular images. The binocular stereo vision has the advantages of high efficiency, high accuracy, simple system structure and low cost. Because the binocular stereo vision needs to match the same point on the left and right image capture points, the focal length and image capture center of the two camera lenses of the camera, as well as a positional relationship between the left and right camera lenses shall be obtained. To obtain the above data, the camera needs to be calibrated. The acquisition of the positional relationship between the visible light camera and the infrared camera is called joint calibration.

In the calibration process, two camera lens parameters and relative position parameters of the camera are obtained, but these parameters are not stable. When temperature and humidity are changed, the internal parameters of the camera lenses are also changed. In addition, due to accidental camera collision, the positional relationship between the two camera lenses may be changed. Therefore, when the camera is used, internal and external parameters must be modified, which is self-calibration. When the internal parameters of the camera are known, the positional relationship between the infrared lens and the visible light lens is corrected by extracting the features of an infrared image and the features of a visible light image respectively, that is, the joint self-calibration of the infrared camera and the visible light camera.

Because the imaging of the infrared camera is different from that of the visible light camera, features with cross-modal robustness shall be found, and the saliency of the image is found as one of the features. The influence of cross-modal imaging differences can be overcome by detecting the saliency of the infrared image and the visible light image and extracting the matching feature points on the saliency images, thereby realizing a joint self-calibration function.

SUMMARY

The present invention aims to solve the change of a positional relationship between an infrared camera and a visible light camera due to factors such as temperature, humidity and vibration. Firstly, the saliency of an infrared image and a visible light image is detected, and matching feature points of an infrared saliency image and a visible light saliency image are extracted to correct an original calibration result.

A multispectral camera dynamic stereo calibration algorithm based on saliency features comprises the following steps:

Step 1: Original image correction: conducting de-distortion and binocular correction on an original image according to internal parameters and original external parameters of the infrared camera and the visible light camera.

Step 2: Detecting the saliency of the infrared image and the visible light image respectively based on a histogram contrast method.

Step 3: Extracting feature points on the infrared image, the visible light image and the saliency images thereof.

Step 4: Matching the feature points extracted in the previous step.

Step 5: Judging a feature point coverage area: dividing the image into m*n grids; if the feature points cover all the grids, executing a next step; otherwise continuing to shoot the image and repeating step 1 to step 4.

Step 6: Correcting the calibration result: using image coordinates of all the feature points to calculate the positional relationship between the two cameras after correction; and then superimposing with the original external parameters.

The step 1 specifically comprises the following steps:

1-1) Calculating the coordinates in a normal coordinate system corresponding to each original image point P_i.

A pixel coordinate system takes the upper left corner of the image as an origin, and x-axis and y-axis of the pixel coordinate system are parallel to x-axis and y-axis of an image coordinate system, respectively. The unit of the pixel coordinate system is the pixel. The pixel is a basic and indivisible unit of image display. The normal coordinate system takes the optical center of the camera as the origin of the image coordinate system and scales the distance from the optical center to an image plane to 1. A relationship between pixel coordinates and normal coordinates is as follows:

$u=\mathrm{KX}\left[\begin{array}{c}u\\ v\\ 1\end{array}\right]=\left[\begin{array}{ccc}{f}_x& 0& c_x\\ 0& f_y& c_y\\ 0& 0& 1\end{array}\right]\left[\begin{array}{c}X\\ Y\\ 1\end{array}\right]$

wherein

$u=\left[\begin{array}{c}u\\ v\\ 1\end{array}\right]$ indicates the pixel coordinate of the image;

$K=\left[\begin{array}{ccc}{f}_x& 0& c_x\\ 0& f_y& c_y\\ 0& 0& 1\end{array}\right]$ indicates an internal parameter matrix of the camera; f_x and f_y respectively indicate the focal distances of the image in x direction and y direction (the unit is the pixel); (c_x, c_y) indicates the principal point position of the camera; and

$X=\left[\begin{array}{c}X\\ Y\\ 1\end{array}\right]$ is a coordinate in the normal coordinate system. The normal coordinate system corresponding to the pixel points can be calculated, i.e.,X=K⁻¹u

The normal coordinate system of each original image point P_i is:X_i=K_iu_i

• • wherein u_i is the pixel coordinate of P_i; X_i is the normal coordinate of P_i; K_i is the internal parameter matrix of the corresponding camera of P_i; if P_i is a point on the infrared image, K_i is the internal parameter matrix of the infrared camera; if P_i is a point on the visible light image, K_i is the internal parameter matrix of the visible light camera.

1-2) Removing image distortion: calculating the normal coordinate of an original image point after de-distortion.

Due to the limitation of a lens production process, a lens under actual conditions has some distortion phenomena, causing nonlinear distortion. The nonlinear distortion can be roughly classified into radial distortion and tangential distortion.

The radial distortion of the image is a position deviation of the image pixel points with the distortion center as the center point along the radial direction, thereby causing the distortion of the picture formed in the image. The radial distortion is roughly described as follows:x_d=x(1+k₁r²+k₂r⁴+k₃r⁶)y_d=y(1+k₁r²+k₂r⁴+k₃r⁶)

• • wherein r²=x²+y²; k₁, k₂ and k₃ are radial distortion parameters.

The tangential distortion is generated by the defect in the camera manufacturing that makes the lens not parallel to the image plane, and can be quantitatively described as:x_d=x+(2p₁xy+p₂(r²+2x²))y_d=y+(p₁(r²+2y²)+2p₂xy)

• • wherein p₁ and p₂ are tangential distortion coefficients.

In conclusion, the coordinate relationship before and after distortion is as follows:x_d=x(1+k₁r²+k₂r⁴+k₃r^b)+(2p₁xy+p₂(r²+2x²))y_d=y(1+k₁r²+k₂r⁴+k₃r⁶)+(p₁(r²+2y²)+2p₂xy)

• • wherein (x,y) is a normal coordinate in an ideal state, and (x_d,y_d) is an actual normal coordinate with distortion. Iterative computation is conducted for several times by taking (x_d, y_d) as the initial value of (x, y) to obtain the actual (x, y).

1-3) Rotating the two images according to the original rotation relationship between the two cameras: an original rotation matrix R and a translation vector t between the two cameras are known:X_r=RX_l+t

• • wherein X_l indicates the normal coordinate of the infrared camera, and X_r indicates the normal coordinate of the visible light camera. The infrared image is rotated to positive direction of R by half an angle, and the visible light image is rotated to opposite direction of R by half an angle.

For the normal coordinate X_i of P_i after de-distortion obtained in the previous step, if P_i is a point on the infrared image, R^1/2 X_i→X_i; if P_i is a point on the visible light image, R^−1/2X_i→X_i.

1-4) Restoring the de-distorted and rotated image to the pixel coordinate system according to the formula u=KX; and calculating the de-distorted and corrected image coordinate according to the normal coordinate P_i of the image point X_i updated in the previous step.K_iX_i→u_i

It can be seen from the above that, if the coordinate u_i of the point before de-distortion and correction is known, the coordinate of the de-distorted and corrected point calculated by steps 1-1)-1-4) is recorded as F(u_i).

1-5) Calculating a pixel coordinate position F⁻¹(v_i) of a corresponding original image I₀ of each image point v_i of the de-distorted and corrected image I, and selecting a color value of the corresponding position from I₀ and filling in I.I(v_i)=I₀(F⁻¹(v_i))

Because F⁻¹(v_i) is a decimal coordinate, bilinear interpolation is needed to calculate the color value of the corresponding position of the decimal coordinate.

The step 2 specifically comprises the following steps:

2-1) Calculating saliency of each color. The saliency of the pixel is determined by the difference between the color of the pixel and the colors of other pixels. Therefore, pixels with the same color have the same saliency value. The saliency value of the pixel color is defined as:

$s\left(I_k\right)=S\left(c_l\right)=\sum _{j=1}^n{f}_{j}D\left(c_l,c_j\right)$

• • wherein n is the total number of the image colors; c_j is the color value of the pixel point I_k; D(c_l, c_j) indicates a distance between two colors c_l and c_j in a color space; and f_j is occurrence probability of color c_j. In order to reduce the number of color types and reduce the calculation complexity, each color channel is quantized into several values.

2-2) Smoothing the color saliency for the purpose of reducing the noise caused by that similar colors are quantified as different values. The formula of the smoothed color saliency is as follows:

$S^{\prime }\left(c\right)=\frac{1}{\left(m-1\right)T}\sum _{i=1}^m\left(T-D\left(c,c_i\right)\right)S\left(c_i\right)$

• • wherein m=n/4; T=Σ_i=1^m D(c, c_i) is the sum of the distances between the color c and m nearest colors.

2-3) Determining the saliency according to the colors on the image pixel points to obtain a saliency image.

The step 3 specifically comprises the following steps:

Difference of Gaussian pyramid is obtained from the difference between adjacent scale spaces, and is often used for scale-invariant feature transform (SIFT). The scale space of an image is defined as: the convolution of a Gaussian convolution kernel and the image, which is a function of parameter σ in the Gaussian convolution kernel. Specifically, the scale space of a scene image I(x, y) is:L(x,y,σ)=G(x,y,σ)*I(x,y)

• • wherein

$G\left(x,y,\sigma \right)=\frac{1}{2\pi \sigma }\times e^{-\frac{x^2+y^2}{2{\sigma }^2}}$ is a Gaussian kernel function: σ is a scale factor; and the value of a determines the smoothness of the image. A large a value corresponds to a rough scale (low resolution), and a small σ value corresponds to a fine scale (high resolution). * indicates convolution operation. L(x,y,σ) is called as the scale space of the image I(x,y). The differences of scale spaces of different scales are calculated to obtain a layer of difference of Gaussian pyramid (as shown in FIG. 3). In addition, a normalized scale factor λ is multiplied so that the maximum value of the DoG image is 255.D(x,y,σ)=λ(L(x,y,kσ)−L(x,y,σ))

Unlike SIFT, only a layer of differential scale features is calculated. Two reasons for this are as follows: firstly, the amount of calculation of multilayer differential scale features is too large, and timeliness cannot be realized; and secondly, the accuracy of SIFT features obtained by using the multilayer differential scale features is too low.

3-2) Taking local extreme points of a single-layer difference of Gaussian pyramid D obtained in the previous step as a feature point set {P}.

3-2-1) Expanding D and recording the result as D₁. Comparing each pixel point in D₁ with a point on an 8-neighborhood; and if the pixel point is a local maximum value, adding the pixel point to a candidate point set P₁.

3-2-2) Negating and expanding D, and recording the result as D₂. Comparing each pixel point in D₂ with a point on an 8-neighborhood; and if the pixel point is a local minimum value, adding the pixel point to a candidate point set P₂.

3-2-3) Intersecting P₁ and P₂ to obtain P₃=P₁∩P₂. Taking points having DoG gray value larger than 15 in P₃ as a feature point set {P}, wherein the feature point set of an infrared gray image is P₀₁^l; the feature point set of an infrared saliency image is P₀₂^l; the feature point set of a visible light gray image is P₀₁^r; and the feature point set of a visible light saliency image is P₀₂^r.

3-2-4) The feature point set of the infrared image is P₀^l=P₀₁^l∩P₀₂^l; and the feature point set of the visible light image is P₀^r=P₀₁^r∩P₀₂^r.

The step 4 specifically comprises the following steps:

4-1) Dividing the infrared image and the visible light image into m×n blocks respectively. Conducting steps 4-2) to 4-6) for each feature point p_i^l of the infrared image.

4-2) Finding a block b_xi,yi^l, corresponding to p_i^l in the infrared image (as shown in FIG. 4(a)), wherein the block at the same position in the visible light image as block b_xi,yi^l is b_xi,yi^r, a block set {B_xi,yi^r} has the same abscissa and ordinate as the block b_xi,yi^r (as shown in FIG. 4(b)), and a feature point set is recorded as {P_i^r}. Assessing the degree of similarity between pixel points. If the degree of similarity is greater than a threshold t₁, regarding the pixel points as rough matching points, and recording the set as {p_i^r}; otherwise, discarding the points and selecting a next feature point to repeat step 4-2).

4-3) If the maximum similarity s_first and the second maximum similarity s_second in p_i^l and {p_i^r} satisfy:F(s_first,s_second)≥t₂

• • reserving the matching, wherein a point p_i^r having the maximum similarity in {p_i^r} is taken as a matching point, t₂ is a threshold and F(s_first, s_second) is used for describing a relationship between s_first and s_second. If the maximum similarity and the second maximum similarity do not satisfy the formula, discarding the points and selecting a next feature point to repeat step 4-2).

After selection according to the rule, matching the corresponding feature point p′_i^l of p_i^r in the infrared image according to the steps 4-2) and 4-3); reserving the matching p_i^l, p_i^r if p′_i^l=p_i^l is satisfied; if p′_i^l=p_i^l is not satisfied, discarding the points and selecting a next feature point to repeat step 4-2).

4-4) Based on the infrared image feature point p_i^l=(x_i^l,y_i^l), conducting parabolic fitting to optimize an integer pixel feature point p_i^r=(x_i^r, y_i^r) corresponding to the visible light image, to obtain a sub-pixel feature point p′_i^r=(x_i^r+j*_rx, y_i^r+j*_ry) corresponding to the visible light image, wherein j*_rx is a sub-pixel offset in x direction and j*_ry is a sub-pixel offset in y direction.

4-5) Based on the integer pixel feature point p_i^r=(x_i^r, y_i^r) corresponding to the visible light image, calculating the sub-pixel feature point p′_i^l=(x_i^l+j*_lx, y_i^r+j*_ly) corresponding to the infrared image according to the method of 4-4), wherein j*_lx is a sub-pixel offset in x direction and j*_ly is a sub-pixel offset in y direction.

4-6) Obtaining a final matching point pair of p′_i^l, p′_i^r. Selecting a next infrared image feature point to repeat steps 4-2) to 4-6).

The step 6 specifically comprises the following steps:

6-1) Solving a basic matrix F and an essential matrix E according to the coordinates of the feature point pairs of the infrared image and the visible light image and the internal parameter matrices of the infrared camera and the visible light camera: a relationship between the pixel point pairs u_l and u_r corresponding to infrared light and visible light and the basic matrix F is:u_r^TFu_l=0

Further screening the point pairs by using random sample consensus (RANSAC); and substituting the coordinates of the corresponding points into the above formula to construct a homogeneous linear equation system to solve F.

A relationship between the basic matrix and the essential matrix is:E=K_r^TFK_l

• • wherein K_l and K_r are respectively the internal parameter matrices of the infrared camera and the visible light camera.

6-2) Decomposing a relationship between rotation and translation of the infrared camera and the visible light camera after correction from the essential matrix: the relationship between the essential matrix E and rotation R and translation t is as follows:E=[t]_xR

• • wherein [t]_x indicates a cross product matrix of t.

Conducting singular value decomposition on E to obtain:

$E=U\sum V^T=U\left[\begin{array}{ccc}1& 0& 0\\ 0& 1& 0\\ 0& 0& 0\end{array}\right]V^T$

Defining two matrices

$Z=\left[\begin{array}{ccc}0& 1& 0\\ -1& 0& 0\\ 0& 0& 0\end{array}\right]\mathrm{and}W=\left[\begin{array}{ccc}0& -1& 0\\ 1& 0& 0\\ 0& 0& 1\end{array}\right],\mathrm{ZW}=\Sigma$

Thus, writing E in the following two forms

• • (1) E=UZU^TUWV^T
  • setting [t]_x=UZU^T, R=UWV^T
  • (2) E=−UZU^TUW^TV^T
  • setting [t]_x=−UZU^T, R=UW^TV^T

Obtaining four pairs of R and t, and selecting a solution with three-dimensional meaning.

6-3) Superimposing the decomposed relationship between rotation and translation into the original external parameters.

Recording the rotation matrix before de-distortion as R₀ and the translation vector as t₀=(t_x, t_y, t_z)^T; recording the rotation matrix calculated in the previous step as R and the translation vector as t=(t′_x, t′_y, t′_x)^T and new R_new and t_new are as follows:

$R_{new}=R_0^{1/2}R{R}_0^{1/2}$ $t_{\mathrm{new}}=R_0^{1/2}{t}_x[\begin{array}{ccc}1& \frac{t_y^{\prime }}{t_x^{\prime }}& \frac{t_z^{\prime }}{t_x^{\prime }}]\end{array}$

Multiplying t_new by a coefficient so that the component of t_new in x direction is t_x^new=t_x.

The present invention has the following beneficial effects:

The present invention solves the change of a positional relationship between an infrared camera and a visible light camera due to factors such as temperature, humidity and vibration. The present invention has the advantages of high speed, accurate results and simple operation. The saliency of the image, which is a feature with cross-modal robustness, is found. The influence of cross-modal imaging differences can be overcome by detecting the saliency of the infrared image and the visible light image and extracting and matching the feature points on the saliency images.

DESCRIPTION OF DRAWINGS

FIG. 1 is an overall flow chart.

FIG. 2 is a correction flow chart.

FIG. 3 shows difference of Gaussian pyramid (DoG).

FIG. 4 is a schematic diagram of block matching.

DETAILED DESCRIPTION

Specific embodiments of the present invention are further described below in combination with accompanying drawings and the technical solution.

Step 1: Original image correction:

1-1) Calculating the coordinates in a normal coordinate system corresponding to each original image point P_i.

The normal coordinate system of each original image point P_i is:X_i=K_iu_i

• • wherein u_i is the pixel coordinate of P_i; X_i is the normal coordinate of P_i; K_i is the internal parameter matrix of the corresponding camera of P_i; if P_i is a point on the infrared image, K_i is the internal parameter matrix of the infrared camera; if P_i is a point on the visible light image, K_i is the internal parameter matrix of the visible light camera.

1-2) Removing image distortion: calculating the normal coordinate of an original image point after de-distortion.

Iterative computation is conducted for several times by taking (x_d, y_d) as the initial value of (x, y) to obtain the actual (x, y).

1-3) Rotating the two images according to the original rotation relationship between the two cameras:

For the normal coordinate X_i of P_i after de-distortion obtained in the previous step, if P_i is a point on the infrared image. R^1/2 X_i→X_i; if P_i is a point on the visible light image, R^−1/2X_i→X_i

1-4) Calculating the de-distorted and corrected image coordinate according to the normal coordinate X_i of the updated image point P_i.K_iX_i→u_i

It can be seen from the above that, if the coordinate u_i of the point before de-distortion and correction is known, the coordinate of the de-distorted and corrected point calculated by steps 1-1)-1-4) is recorded as F(u_i).

1-5) Calculating a pixel coordinate position F⁻¹(v_i) of a corresponding original image I₀ of each image point v_i of the de-distorted and corrected image I, and selecting a color value of the corresponding position from I₀ and filling in I:I(v_i)=I₀(F⁻¹(v_i))

Because F⁻¹(v_i) is a decimal coordinate, bilinear interpolation is needed to calculate the color value of the corresponding position of the decimal coordinate.

2) Detecting the saliency of the infrared image and the visible light image respectively based on a histogram contrast method.

2-1) Calculating saliency of each color. The saliency of the pixel is determined by the difference between the color of the pixel and the colors of other pixels. Therefore, pixels with the same color have the same saliency value. The saliency value of the pixel color is defined as:

$S\left(I_k\right)=S\left(c_i\right)=\sum _{j=1}^n{f}_{j}D\left(c_l,c_j\right)$

• • wherein n is the total number of the image colors; c_j is e color value of the pixel point I_k: D(c_l, c_j) indicates a distance between two colors c_l and c_j in a color space; and f_j is occurrence probability of color c_j. In order to reduce the number of color types and reduce the calculation complexity, each color channel is quantized into several values.

2-2) Smoothing the color saliency for the purpose of reducing the noise caused by that similar colors are quantified as different values. The formula of the smoothed color saliency is as follows:

$S^{\prime }\left(c\right)=\frac{1}{\left(m-1\right)T}\sum _{i=1}^m\left(T-D\left(c,c_i\right)\right)S\left(c_i\right)$

• • wherein m=n/4; T=Σ_i=1^m D(c, c_i) is the sum of the distances between the color c and m nearest colors.

2-3) Determining the saliency according to the colors on the image pixel points to obtain a saliency image.

3) Extracting feature points on the infrared image, the visible light image and the saliency images thereof.

3-1) Constructing a multi-scale difference of Gaussian pyramid (DoG). The DoG pyramid is obtained from the difference between adjacent scale spaces, and is often used for scale-invariant feature transform (SIFT). The scale space of an image is defined as: the convolution of a Gaussian convolution kernel and the image, which is a function of parameter σ in the Gaussian convolution kernel. Specifically, the scale space of a scene image I(x,y) is:L(x,y,σ)=G(x,y,σ)*I(x,y)

wherein

$G\left(x,y,\sigma \right)=\frac{1}{2\pi {\sigma }^2}\times e^{-\frac{x^2+y^2}{2{\sigma }^2}}$ is a Gaussian kernel function: σ is a scale factor; and the value of a determines the smoothness of the image. A large σ value corresponds to a rough scale (low resolution), and a small σ value corresponds to a fine scale (high resolution). * indicates convolution operation. L(x,y,σ) is called as the scale space of the image I(x,y). The differences of scale spaces of different scales are calculated to obtain a layer of difference of Gaussian pyramid (as shown in FIG. 3). In addition, a normalized scale factor λ is multiplied so that the maximum value of the DoG image is 255.D(x,y,σ)=λ(L(x,y,kσ)−L(x,y,σ))

Unlike SIFT, only a layer of differential scale features is calculated. Two reasons for this are as follows; firstly, the amount of calculation of multilayer differential scale features is too large, and timeliness cannot be realized; and secondly, the accuracy of SIFT features obtained by using the multilayer differential scale features is too low.

3-2) Comparing each point in the DoG obtained in the previous step with the pixel points in the neighborhood to judge whether the point is a local extreme point.

3-2-1) Expanding D and recording the result as D₁. Comparing each pixel point in D₁ with a point on an 8-neighborhood; and if the pixel point is a local maximum value, adding the pixel point to a candidate point set P₁.

3-2-2) Negating and expanding D, and recording the result as D₂. Comparing each pixel point in D₂ with a point on an 8-neighborhood; and if the pixel point is a local minimum value, adding the pixel point to a candidate point set P₂.

3-2-3) Intersecting P₁ and P₂ to obtain P₃=P₁∩P₂. Taking points having DoG gray value larger than 15 in P₃ as a feature point set {P}, wherein the feature point set of an infrared gray image is P₀₁^l; the feature point set of an infrared saliency image is P₀₂^l; the feature point set of a visible light gray image is P₀₁^r; and the feature point set of a visible light saliency image is P₀₂^r.

3-2-4) The feature point set of the infrared image is P₀^l=P₀₁^l∩P₀₂^l; and the feature point set of the visible light image is P₀^r=P₀₁^r∩P₀₂^r.

4) Matching the feature points extracted in the previous step.

4-1) Dividing the infrared image and the visible light image into m×n blocks respectively; and conducting steps 4-2) to 4-6) for each corner point p_i^l of the infrared image.

4-2) Finding a block b_xi,yi^l corresponding to p_i^l in the infrared image, wherein the block at the same position in the visible light image as block b_xi,yi^l is b_xi,yi^r, a block set {B_xi,yi^r} has the same abscissa and ordinate as the block b_xi,yi^r and a corner point set is recorded as {P_l^r}.

Calculating the degree of similarity of any point in p_i^l and {P_i^r}. If the degree of similarity is greater than a threshold t₁, regarding the point as rough matching point, and recording the set as {p_i^r}; otherwise, discarding the point and selecting a next corner point to repeat step 4-2).

4-3) If the maximum similarity s_first and the second maximum similarity s_second in p_i^l and {p_i^r} satisfy:F(s_first,s_second)≥t₂

Reserving the matching, wherein a point p_i^r having the maximum similarity in {p_i^r} is taken as a matching point, t₂ is a threshold and F(s_first, s_second) is used for describing a relationship between s_first and s_second; if the maximum similarity and the second maximum similarity do not satisfy the formula, discarding the point and selecting a next corner point to repeat step 4-2).

After selection according to the rule, matching the corresponding corner point p′_i^l of p_i^r in the infrared image according to the steps 4-2) and 4-3); reserving the matching p_i^l, p_i^r if p′_i^l=p_i^l is satisfied; if p′_i^l=p_i^l is not satisfied, discarding the points and selecting a next feature point to repeat step 4-2).

4-4) Based on the infrared image corner point p_i^l=(x_i^l, y_i^l), conducting parabolic fitting to optimize an integer pixel corner point p_i^r=(x_i^r, y_i^r) corresponding to the visible light image, to obtain a sub-pixel corner point p′_i^r=(x_i^r+j*_rx, y_i^r+j*_ry) corresponding to the visible light image, wherein j*_rx is a sub-pixel offset in x direction and j*_ry is a sub-pixel offset in y direction.

4-5) Based on the integer pixel corner point p_i^r=(x_i^r,y_i^r) corresponding to the visible light image, calculating the sub-pixel corner point p′_i^l=(x_i^l+j*_lx, y_i^r+j*_ly) corresponding to the infrared image according to the method of 4-4), wherein j*_lx is a sub-pixel offset in x direction and j*_ly is a sub-pixel offset in y direction.

4-6) Obtaining a final matching point pair of p′_i^l, p′_i^r ; and selecting a next infrared image corner point to repeat steps 4-2) to 4-6).

5) Judging a feature point coverage area: dividing the image into m*n grids; if the feature points cover all the grids, executing a next step; otherwise continuing to shoot the image and repeating step 1) to step 4).

6) Correcting the calibration result: using image coordinates of all the feature points to calculate the positional relationship between the two cameras after correction; and then superimposing with the original external parameters.

6-1) Further screening the point pairs by using random sample consensus (RANSAC); and substituting the coordinates of the corresponding points into the above formula to construct a homogeneous linear equation system to solve F.

A relationship between the basic matrix and the essential matrix is:E=K_r^TFK_l

• • wherein K_l and K_r are respectively the internal parameter matrices of the left camera and the right camera.

6-2) Decomposing a relationship between rotation and translation of the left camera and the right camera after correction from the essential matrix: the relationship between the essential matrix E and rotation R and translation t is as follows:E=[t]_xR

• • wherein [t]_x indicates a cross product matrix of t;

Conducting singular value decomposition on E to obtain:

$E={\mathrm{U\Sigma V}}^T=U\left[\begin{array}{ccc}1& 0& 0\\ 0& 1& 0\\ 0& 0& 0\end{array}\right]V^T$

Defining two matrices

$Z=\left[\begin{array}{ccc}0& 1& 0\\ -1& 0& 0\\ 0& 0& 0\end{array}\right]\mathrm{and}W=\left[\begin{array}{ccc}0& -1& 0\\ 1& 0& 0\\ 0& 0& 1\end{array}\right],\mathrm{ZW}=\Sigma$

Thus, writing E in the following two forms

• • (1) E=UZU^TUWV^T
  • setting [t], =UZU^T, R=UWV^T
  • (2) E=−UZU^TUW^TV^T
  • setting [t]_x=−UZU^T, R=UW^TV^T

Obtaining four pairs of R and t, and selecting a solution with three-dimensional meaning.

6-3) Superimposing the decomposed relationship between rotation and translation into the original external parameters.

Calculation results.

The rotation matrix R and the translation vector t before de-distortion are:

$R=\left[\begin{array}{ccc}0.999990& -0.004293& 0.000874\\ 0.004295& 0.999987& -0.002599\\ -0.000863& 0.002603& 0.999996\end{array}\right]$ $t=[\begin{array}{ccc}-342.4272& -2.0244& {-0.7514]}^T\end{array}$

The calculated rotation matrix is R′, and the translation vector is t′

$R^{\prime }=\left[\begin{array}{ccc}0.999999& -0.001413& 0.000990\\ 0.001413& 0.999999& -0.000238\\ -0.000990& 0.000239& 0.999999\end{array}\right]$ $t^{\prime }=[\begin{array}{ccc}-1.000& -0.0050& {-0.0075]}^T\end{array}$

New R_new and t_new are:

$R_{\mathrm{new}}=\left[\begin{array}{ccc}0.999982& -0.005704& 0.001866\\ 0.005709& 0.999980& -0.002834\\ -0.001850& 0.002845& 0.999994\end{array}\right]$ $t_{\mathrm{new}}=[\begin{array}{ccc}-342.4272& 0.9829& {-2.4117]}^T.\end{array}$

Claims

1. A multispectral camera dynamic stereo calibration algorithm based on saliency features, stored on a non-transitory computer-readable medium, and when executed by a processor, causes the processor to perform the following steps: step 1: original image correction: conducting de-distortion and binocular correction on an original image according to internal parameters and original external parameters of an infrared camera and a visible light camera;step 2: detecting a saliency of an infrared image and a visible light image respectively based on a histogram contrast method;2-1) calculating saliency of each color; and defining a saliency value of the pixel color as:

2. The multispectral camera dynamic stereo calibration algorithm based on saliency features according to claim 1, wherein the step 1 comprises the following steps:1-1) calculating the image coordinates in a normal coordinate system corresponding to each original image point Pi;the normal coordinate system of each original image point Pi is: Xi=Kiui wherein ui is a pixel coordinate of Pi; Xi is a normal coordinate of Pi; Ki is an internal parameter matrix of a corresponding infrared camera of Pi; if Pi is a point on the infrared image, Ki is the internal parameter matrix of the infrared camera; if Pi is a point on the visible light image, Ki is the internal parameter matrix of the visible light camera;1-2) removing image distortion: calculating the normal coordinate of an original image point after de-distortion;iterative computation is conducted for several times by taking (xd, yd) as an initial value of (x, y) to obtain the actual (x, y);1-3) rotating the infrared image and the visible light image according to the original rotation relationship between the infrared image camera and the visible light image camera: an original rotation matrix R and a translation vector t between the infrared image camera and the visible light image camera are known: Xr=RXl+t wherein Xl indicates the normal coordinate of the infrared camera, and Xr indicates the normal coordinate of the visible light camera; the infrared image is rotated to positive direction of R by half an angle, and the visible light image is rotated to opposite direction of R by half an angle;for the normal coordinate Xi of Pi after de-distortion obtained in the previous step, if Pi is a point on the infrared image, R1/2Xi→Xl; if Pi is a point on the visible light image, R−1/2Xi→Xi 1-4) restoring the de-distorted and rotated image to the pixel coordinate system according to the formula u=KX; and calculating the de-distorted and corrected image coordinate according to the normal coordinate Xi of the image point Pi updated in the previous step; KiXi→ui if the coordinate ui of the point before de-distortion and correction is known, the coordinate of the de-distorted and corrected point calculated by steps 1-1)-1-4) is recorded as F(ui);1-5) calculating a pixel coordinate position F−1(vi) of a corresponding original image I0 of each image point vi of a de-distorted and corrected image I, and selecting a color value of a corresponding position from I0 and filling in I: I(vi)=I0(F−1(vi))because F−1(vi) is a decimal coordinate, bilinear interpolation is needed to calculate the color value of the corresponding position of the decimal coordinate.

3. The multispectral camera dynamic stereo calibration algorithm based on saliency features according to claim 1, wherein the step 6 comprises the following steps: 6-1) solving a basic matrix F and an essential matrix E according to the coordinates of the feature point pairs of the infrared image and the visible light image and the internal parameter matrices of the infrared camera and the visible light camera: a relationship between the pixel point pairs ui and ur corresponding to infrared light and visible light and the basic matrix F is: urTFul=0further screening the point pairs by using random sample consensus (RANSAC); and substituting the coordinates of the corresponding points into the above formula to construct a homogeneous linear equation system to solve the basic matrix F;a relationship between the basic matrix and the essential matrix is: E=KrTFKl wherein Kl and Kr are respectively the internal parameter matrices of the infrared camera and the visible light camera;6-2) decomposing a relationship between rotation and translation of the infrared camera and the visible light camera after correction from the essential matrix: the relationship between the essential matrix E and rotation R and translation t is as follows: E=[t]xR wherein [t]x indicates a cross product matrix of t;conducting singular value decomposition on E to obtain:

4. The multispectral camera dynamic stereo calibration algorithm based on saliency features according to claim 3, wherein the step 3 comprises the following steps: taking local extreme points of an obtained single-layer difference of Gaussian pyramid D as a feature point set {P};3-2-1) expanding the Gaussian pyramid D and recording a result as D1; comparing each pixel point in the result D1 with a point on an 8-neighborhood; and if the pixel point is a local maximum value, adding the pixel point to a candidate point set P1;3-2-2) negating and expanding the Gaussian pyramid D, and recording the result as D2;comparing each pixel point in the result D2 with a point on an 8-neighborhood; and if the pixel point is a local minimum value, adding the pixel point to a candidate point set P2;3-2-3) intersecting the candidate set points P1 and P2 to obtain a set point P3=P1∩P2; taking points having DoG gray value larger than 15 in the set point P3 as a feature point set {P}, wherein the feature point set of an infrared gray image is P01l; the feature point set of an infrared saliency image is P02l; the feature point set of a visible light gray image is P01r; and the feature point set of a visible light saliency image is P02r;3-2-4) the feature point set of the infrared image is P0l=P01l∪P02l; and the feature point set of the visible light image is P0r=P01r∪P02r.

5. The multispectral camera dynamic stereo calibration algorithm based on saliency features according to claim 1, wherein the step 3 comprises the following steps: taking local extreme points of an obtained single-layer difference of Gaussian pyramid D as a feature point set {P};3-2-1) expanding the Gaussian pyramid D and recording a result as D1; comparing each pixel point in the result D1 with a point on an 8-neighborhood; and if the pixel point is a local maximum value, adding the pixel point to a candidate point set P1;3-2-2) negating and expanding the Gaussian pyramid D, and recording a result as D2; comparing each pixel point in the result D2 with a point on an 8-neighborhood; and if the pixel point is a local minimum value, adding the pixel point to a candidate point set P2;3-2-3) intersecting the candidate point sets P1 and P2 to obtain a point set P3=P1 ∩P2; taking points having DoG gray value larger than 15 in the point set P3 as a feature point set {P}, wherein the feature point set of an infrared gray image is P01l; the feature point set of an infrared saliency image is P02l; the feature point set of a visible light gray image is P01r; and the feature point set of a visible light saliency image is P02r;3-2-4) the feature point set of the infrared image is P0l=P01l∪P02l; and the feature point set of the visible light image is P0r=P01r∪P02r.