Long-distance binocular camera calibration optimization method based on multistage feature enhancement

Abstract

The invention discloses a long-distance binocular camera calibration optimization method based on multistage feature enhancement, and the method comprises the steps: carrying out the processing of a condition that a long-distance calibration plate has a highlight region and is fuzzy in angular points, employing a bilateral filter to suppress the reflection of a highlight mirror surface, reducing the reflection of light, and maintaining the definition of an image; the checkerboard texture is enhanced by adopting a CLAHE algorithm in combination with blocking processing and a contrast gain threshold value; a Sobel operator and a Harris corner response function are fused, gradient direction distribution characteristics of a pixel neighborhood are extracted through the Sobel operator, weighted fusion is carried out on the gradient direction distribution characteristics and the Harris corner response function, and the response intensity and specificity of a corner area are remarkably enhanced. A three-level image preprocessing framework including reflection suppression, contrast enhancement and corner enhancement is constructed, the problem of feature extraction of a traditional method under a complex illumination condition is solved, the corner area detection capability is enhanced, the problems of specular reflection noise, low contrast and corner blur are effectively solved, and the calibration precision is improved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of computer vision and optical engineering, in particular to a long-distance binocular camera calibration optimization method based on multi-level feature enhancement. BACKGROUND

[0002] In the prior art, there are common limitations in complex environments, which are poor adaptability to light changes and dynamic interference, and lack of multi-level enhancement mechanism for feature extraction, resulting in significant decline in calibration accuracy. This is embodied in the following aspects: for the processing of light reflection interference, it is mostly dependent on hardware devices such as polarization filters, or complex post-processing techniques such as RealSense TM encoding light technology in SR300, but lacks a lightweight solution based on image enhancement; traditional methods rely on global histogram equalization in low-contrast environments, but are prone to cause local texture over-enhancement and noise amplification, resulting in a decline in checkerboard feature extraction accuracy; and multi-camera systems are mostly used to improve feature robustness, but the multi-level gradient information of monocular images is not fully exploited. In long-distance calibration, due to light scattering and resolution limitations, the checkerboard corner points are blurred, making it difficult for traditional algorithms to accurately locate them, which significantly reduces the calibration accuracy. SUMMARY

[0003] The present disclosure aims to provide a long-distance binocular camera calibration optimization method based on multi-level feature enhancement, which strengthens the corner region detection capability, effectively solves the problems of mirror reflection noise, low contrast and corner point blur, and improves the calibration accuracy.

[0004] According to one of the schemes of the present disclosure, the long-distance binocular camera calibration optimization method based on multi-level feature enhancement comprises:

[0005] The high light area of the calibration board is processed to suppress mirror reflection noise while retaining image edge features, balancing smoothing processing and edge information retention;

[0006] CLAHE algorithm is used in combination with block processing and contrast gain threshold to enhance the checkerboard texture;

[0007] A bilateral filter is used to suppress high light mirror reflection and strengthen the image corner points, and the CLAHE algorithm is used to optimize the image gray scale distribution to enhance the local contrast features of the checkerboard texture.

[0008] In some embodiments, wherein the high light area of the calibration board is processed, comprising:

[0009] A bilateral filter is used to process the image, and the weight of the bilateral filter constructs a fusion space with dual dimensions of space and brightness;

[0010] Through the multiplication operation of the spatial and brightness weights, the bilateral filtering retains the edge pixels based on the brightness mutation characteristics of the edge region when suppressing the specular reflection noise.

[0011] In some embodiments, the spatial weight is calculated by the formula

[0012]

[0013] The metric pixel position correlation is associated, where σ s The influence of the spatial distance on the weight is regulated to ensure the spatial correlation constraint of the neighborhood pixels.

[0014] The brightness weight is calculated by the formula

[0015]

[0016] The metric pixel gray difference is carved, where σ r The influence of the brightness change on the weight is controlled to make the algorithm sensitive to the abnormal brightness change caused by the reflection, and the specular reflection interference is specifically suppressed.

[0017] In some embodiments, the filtering formula is:

[0018]

[0019] Wherein Ω is the neighborhood window centered on x;

[0020] W total =∑ y∈Ω W s (x,y)·W r (x,y).

[0021] In some embodiments, the CLAHE algorithm is combined with the block processing and the contrast gain threshold to enhance the checkerboard texture, including:

[0022] The image is divided into grid sub-regions of a preset size, and the geometric center of each sub-region is taken as a control point;

[0023] The spatial coordinates and head and tail data of the sub-regions are independently processed and allocated;

[0024] The histogram is truncated by dynamic range limitation, and the pixel intensity difference of the truncated part is uniformly redistributed in the sub-region. This process may again exceed the threshold to trigger secondary overflow. Through subsequent iterative optimization processing, the overflow amount is lower than the preset tolerance threshold, and the histogram form is ensured to be stable and convergent.

[0025] Finally, the adaptive enhancement of the local contrast of the checkerboard texture is realized, and the output result is obtained.

[0026] In some embodiments, the method further comprises:

[0027] The method of fusing Sobel operator and Harris corner response function extracts the gradient direction distribution characteristics of pixel neighborhood, fuses with Harris corner response function, enhances the response value of corner region, and realizes robust feature enhancement.

[0028] In some embodiments, the Sobel-Harris fusion enhanced corner method comprises:

[0029] Gradient calculation to obtain gradient amplitude;

[0030] Constructing local structure tensor and performing scale space integration by Gaussian smoothing;

[0031] Harris corner response calculation;

[0032] Feature fusion and contrast enhancement use gradient amplitude features and are processed by normalized linear fusion.

[0033] In some embodiments, obtaining the gradient amplitude comprises:

[0034] Sobel operator convolution is performed on the input gray image I(x, y) to calculate the horizontal gradient field G x and the vertical gradient field G y .

[0035] Based on the horizontal gradient field G x and the vertical gradient field G y , the gradient amplitude

[0036] In some embodiments, constructing the local structure tensor and performing scale space integration by Gaussian smoothing comprises:

[0037] The matrix components and G xy = G x ⊙G y , wherein ⊙ represents element-wise multiplication;

[0038] The Gaussian window size is

[0039] In some embodiments, feature fusion and contrast enhancement use gradient amplitude features and are processed by normalized linear fusion, comprising:

[0040] A coefficient β is introduced to determine a best value, to produce local brightness attenuation in the high value area, enhance the edge contrast, and keep the original gray distribution in the low value area to suppress noise amplification.

[0041] The method of various embodiments of the present disclosure at least processes the high light area existing on the calibration board, retains the image edge features while suppressing the specular reflection noise, balances the smoothing processing and the edge information retention; adopts the CLAHE algorithm combined with the blocking processing and the contrast gain threshold to enhance the checkerboard texture; adopts the bilateral filter to suppress the high light specular reflection, strengthens the image corner points, optimizes the image gray scale distribution combined with the CLAHE algorithm to enhance the local contrast feature of the checkerboard texture, thereby proposing a multi-level feature enhancement long-distance binocular camera calibration optimization method, suppresses the high light reflection and retains the edge sharpness through the bilateral filtering, enhances the texture saliency in the low light area combined with the CLAHE algorithm, and innovatively fuses the Sobel gradient direction and the Harris corner point response function to strengthen the corner point area detection capability, effectively solves the problems of specular reflection noise, low contrast and corner point blur, and improves the calibration accuracy.

[0042] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure as claimed. BRIEF DESCRIPTION OF DRAWINGS

[0043] In the drawings, which are not necessarily drawn to scale, like numerals can describe similar components in different views. Like numerals having different letter suffixes can represent different instances of similar components. The drawings illustrate generally, by way of example, various embodiments in which the disclosure can be practiced as disclosed. In the drawings:

[0044] Figure 1 An algorithm flowchart for an embodiment of the present disclosure;

[0045] Figure 2 A checkerboard calibration image group captured by a left camera of an embodiment of the present disclosure;

[0046] Figure 3 A checkerboard calibration image group captured by a right camera of an embodiment of the present disclosure. DETAILED DESCRIPTION

[0047] In order to make the purposes, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions of the embodiments of the present disclosure will be described clearly and completely below in conjunction with the drawings of the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, rather than all the embodiments. Based on the described embodiments of the present disclosure, all other embodiments obtained by a person of ordinary skill in the art without any inventive effort fall within the protection scope of the present disclosure.

[0048] As the core of three-dimensional perception system, the accuracy of binocular vision calibration has a decisive role in the reliability of key tasks such as stereo matching and depth estimation. However, in actual application scenarios, traditional calibration methods face many challenges in complex scenes.

[0049] Under dynamic lighting conditions, traditional chessboard corner recognition methods are easily disturbed by mirror reflection, leading to positioning errors. Although some methods have improved the feature extraction accuracy to some extent, due to the complex dynamic characteristics of reflection interference, their scope of application is still limited. In addition, methods such as using laser radar to scan three-dimensional calibration targets to achieve accurate geometric calibration rely on high-precision equipment and are difficult to be widely applied to ordinary chessboard calibration scenes.

[0050] In low-contrast environments such as overcast or backlight scenes, the edge saliency of the chessboard is weakened, and traditional corner detection algorithms cause feature detection or false detection due to reduced response. Data-driven methods are more sensitive to initial parameters in low-contrast scenes, further exacerbating calibration errors. Point-line fusion methods enhance matching robustness through gradient direction distribution features, but do not optimize local contrast under low-contrast conditions.

[0051] Dynamic blur can also affect traditional calibration methods, causing traditional methods that rely on stable corner matching to fail. Geometric appearance fusion strategies are used to correct calibration errors, but this research mainly focuses on multi-view human pose estimation, and further exploration of feature enhancement mechanisms in long-distance dual target calibration is still needed.

[0052] In combination with the content described in the background technology section, the present disclosure exemplarily describes the corresponding solutions in the embodiments to solve the defects in the prior art, but not as a limitation on the scope of patent protection required by the present disclosure.

[0053] As one of the solutions, the embodiments of the present disclosure provide a long-distance binocular camera calibration optimization method based on multi-level feature enhancement, comprising:

[0054] Processing the highlight area of the calibration board to suppress mirror reflection noise while preserving image edge features, balancing smoothing processing and edge information retention;

[0055] CLAHE algorithm is used in combination with block processing and contrast gain threshold to enhance the chessboard texture;

[0056] A bilateral filter is used to suppress high light mirror reflection and strengthen image corners, and the CLAHE algorithm is used to optimize image gray scale distribution to enhance the local contrast features of the chessboard texture.

[0057] In combination with the foregoing, to solve the above problems, one of the concepts of the present disclosure is at least to propose a multi-level feature enhancement remote binocular camera calibration optimization method, a multi-level feature enhancement processing framework is constructed, which includes reflection suppression, contrast enhancement and corner point strengthening, and the feature extraction problem of the traditional chessboard calibration method under complex lighting conditions is solved. In addition, a Sobel-Harris fusion enhanced corner point intensity method is proposed, which effectively improves the robustness of feature extraction of the calibration board while maintaining the integrity of the image geometric structure, and significantly enhances the clarity of the remote chessboard corner points. The multi-level preprocessing framework constructed can reduce the re-projection error of calibration and improve the calibration accuracy. After using the Sobel-Harris fusion enhanced corner point intensity method, the re-projection error is further reduced.

[0058] The steps of the embodiments of the present disclosure are exemplary, and the multi-level feature enhancement remote binocular camera calibration optimization method of the present disclosure is further described below by taking steps S1 to S3 as an example.

[0059] In some specific solutions, the embodiments of the present disclosure can be: processing the highlight area of the calibration board, including:

[0060] The image is processed using a bilateral filter, and the weight of the bilateral filter constructs a fusion space and a brightness double dimension;

[0061] Through the product operation of the space and brightness weight, the bilateral filter retains the edge pixels based on the brightness mutation characteristics of the edge area when suppressing the mirror reflection noise.

[0062] Step S1: processing the highlight area of the calibration board.

[0063] First, the image is processed using a bilateral filter, which retains the image edge features while suppressing mirror reflection noise, balances smoothing processing and edge information retention, and provides an effective solution to the image clarity problem under reflection interference.

[0064] The weight of the bilateral filter constructs a fusion space and a brightness double dimension.

[0065] Spatial weight: through the formula

[0066]

[0067] Measure the correlation of pixel positions, where σ s Adjust the influence of spatial distance on weight, ||x-y|| represents the Euclidean distance between coordinates x and y, measures the interval of two pixels in spatial position, and ensures the spatial correlation constraint of adjacent pixels.

[0068] Brightness weight: use the formula

[0069]

[0070] Image gray difference, σ r Control the effect of brightness change on weight, I(x) is the pixel brightness value in the horizontal direction, I(y) is the pixel brightness value in the vertical direction, make the algorithm sensitive to abnormal brightness change caused by reflection, and specifically suppress the interference of specular reflection. The final filtering formula is:

[0071]

[0072] Ω is a neighborhood window centered on x;

[0073] W total =∑ y∈Ω W s (x,y)·W r (x,y) represents the total weight.

[0074] Through the multiplication operation of spatial and brightness weight, the bilateral filter can retain edge pixels based on the brightness mutation characteristics of the edge region when suppressing specular reflection noise, realizing the cooperative optimization of reflection suppression and edge preservation.

[0075] To verify the advantages of the bilateral filter in suppressing specular reflection and preserving edge gradient, as shown in the left checkerboard image in Fig. Figure 2 , 15 photos taken by the left camera were used for experiments, and the bilateral filter was compared with the Gaussian filter and the median filter, as shown in Table 1. According to Table 1, the bilateral filter shows a significant advantage in edge gradient preservation rate, with an average EGPR (Edge Gradient Preservation Rate) of 97.49% and a standard deviation of only 0.506%, which is much higher than the Gaussian filter and the median filter. This advantage is due to the dual weighting mechanism of the spatial domain and the gray domain: the spatial domain standard deviation effectively suppresses noise and preserves edge sharpness, and the gray domain standard deviation distinguishes high light interference from real texture, thereby realizing robust feature preservation under complex lighting.

[0076] Table 1 Comparison of edge gradient preservation performance of filters

[0077] Filter type Mean EGPR (%) Standard deviation (%) Gaussian filter 75.6968 1.8014 Median filter 83.9116 1.2567 Bilateral filter 97.4863 0.5060

[0078] In some specific schemes, the embodiments of the present disclosure can be: using CLAHE algorithm combined with block processing and contrast gain threshold to enhance the checkerboard texture, including:

[0079] The image is divided into grid sub-regions of a preset size, with the geometric center of each sub-region serving as the control point. The spatial coordinates and head-tail data allocation of each sub-region are processed independently. The histogram is truncated through dynamic range limiting, and the pixel intensity difference of the truncated part is evenly redistributed to the sub-regions. This process may exceed the threshold again, causing secondary overflow. Subsequent iterations are used to optimize the process until the overflow amount is lower than the preset tolerance threshold, ensuring that the histogram shape is stable and convergent. Finally, adaptive enhancement of local contrast of the checkerboard texture is achieved, and the result is output.

[0080] Step S2: Enhance the checkerboard texture.

[0081] like Figure 1 The algorithm flowchart shown illustrates the CLAHE algorithm combined with block processing and contrast gain thresholding for addressing the edge blurring problem in outdoor calibration checkerboard patterns. The image is uniformly divided into 8×8 pixel non-overlapping rectangular sub-regions, with the geometric center of each sub-region serving as the reference grid point. Gray-level peak values ​​are independently calculated and redistributed for each sub-region, and the histogram is truncated at the top using a predefined contrast gain threshold. The truncated pixel values ​​are then uniformly redistributed to all gray levels in the current sub-region; this process may trigger secondary gray-level overflow. The overflow amount is checked to ensure it is below a preset threshold; if not, the "truncation-redistribution" process is iteratively executed until the overflow amount falls below the preset threshold, ensuring stable convergence of the histogram shape. Ultimately, this improves the local contrast of the checkerboard texture, effectively mitigating the edge blurring problem.

[0082] In some specific solutions, the embodiments of this disclosure can be: a method that integrates the Sobel operator and the Harris corner response function, by extracting the gradient direction distribution features of the pixel neighborhood and integrating them with the Harris corner response function, thereby enhancing the response value of the corner region and achieving robust feature enhancement.

[0083] Step S3: Enhance the corner points of the image.

[0084] To address the feature blurring issue in camera calibration, a bilateral filter is proposed to suppress specular reflections, and the CLAHE algorithm is combined to optimize image grayscale distribution, thereby enhancing the local contrast features of the checkerboard texture. Building upon this, a method fusing the Sobel operator and the Harris corner response function is proposed. By extracting the gradient direction distribution features of the pixel neighborhood and fusing them with the Harris corner response function, the response values ​​of the corner regions are enhanced, achieving robust feature enhancement.

[0085] The Sobel-Harris fusion enhancement corner point method flow is as follows:

[0086] Gradient calculation: Perform Sobel convolution on the input grayscale image I(x,y), S x For the horizontal Sobel operator convolution kernel, Sy is the Sobel operator kernel in vertical direction, calculating the horizontal gradient field G x is the Sobel operator kernel in vertical direction, calculating the horizontal gradient field G y :

[0087]

[0088] is the gradient magnitude

[0089]

[0090] Construct local structure tensor: matrix components and G xy = G x ⊙ G y , where ⊙ denotes element-wise multiplication. Matrix components is the smoothed horizontal gradient square, reflecting the stability of horizontal edges in local region, is the smoothed vertical gradient square, reflecting the stability of vertical edges in local region.

[0091] Subsequently, scale space integration is performed using Gaussian smoothing:

[0092]

[0093] where the Gaussian window size is set as size is the Gaussian window size, and σ is the standard deviation of Gaussian function.

[0094] Harris corner response calculation: based on the smoothed tensor components, M is the structure tensor matrix, and the improved Harris corner response is calculated:

[0095]

[0096] where det(M) represents the corner strength, and the larger the value, the more significant the corner; trace 2 (M) represents the edge strength, and k controls the edge suppression strength.

[0097] Feature fusion and contrast enhancement use the Sobel gradient magnitude feature calculated in the previous formula Through normalized linear fusion processing:

[0098]

[0099] Balancing the balance between corner specificity and edge continuity.

[0100] where R max is the maximum value of R Harris in all points in the image, and R min is the RHarris The minimum value; It is the gradient magnitude of all points in the image. Maximum value It is the gradient magnitude of all points in the image. Minimum value. α is a weighting coefficient used to weigh R. Harris The greater the weight of image gradient magnitude in the fusion calculation, the greater R. Harris The higher the proportion in the fusion result, the better. α∈[0.6,0.8] controls the feature weights.

[0101] Finally, non-linear contrast enhancement is performed:

[0102] I enhanced =mat2gray(I-β·R) fused )

[0103] In the formula, mat2gray is a MATLAB function that linearly maps a matrix to a grayscale range of [0,1]. The coefficient β can be determined through multiple experiments to find the optimal value. Its mechanism of action is as follows: In R... fused High-value areas, i.e., corners, experience local brightness attenuation, enhancing edge contrast; while low-value areas maintain the original grayscale distribution, suppressing noise amplification. This embodiment's entire method employs a chain-like process of "gradient field construction → tensor modeling → feature fusion → contrast enhancement," achieving robust corner feature enhancement in long-distance calibration scenarios and effectively improving the reliability of corner detection.

[0104] The technical solutions and beneficial effects of the various embodiments of this disclosure are illustrated below with reference to the accompanying drawings and some specific implementation methods.

[0105] Example 1 is used as an example for illustration:

[0106] In this embodiment, a binocular camera is used for image acquisition, outputting raw images with a resolution of 5488 pixels × 3672 pixels. Under natural lighting conditions, multi-view data acquisition is performed on a 12×9 checkerboard calibration board located at a distance outdoors. Each checkerboard square measures 50.00 mm × 50.00 mm. Fifteen sets of stereo vision datasets of the calibration board's pose are collected at different distances and angles. The images acquired by the camera are shown below. Figure 2 and Figure 3 As shown, where Figure 2 Taken by the camera on the left. Figure 3 This photo was taken by the camera on the right.

[0107] After processing the acquired images using the following three sets of methods, calibration is performed.

[0108] Group 1: Use the raw images directly without any preprocessing.

[0109] The second group: first, the image is processed by bilateral filtering to suppress the reflection effect in the high light area and maintain the image edge sharpness. Then, the image gray distribution is optimized by CLAHE algorithm. Finally, the traditional Harris corner response function is used to directly enhance the corner points of the checkerboard image processed above.

[0110] The third group: the first two steps are the same as the second group experiment, bilateral filtering is performed first, and then CLAHE processing is performed. Finally, the Sobel-Harris fusion algorithm is used to enhance the corner response of the checkerboard, which is the method of the embodiments of the present disclosure.

[0111] The calibration results are shown in Table 2 below.

[0112] Table 2 Calibration results

[0113]

[0114]

[0115]

[0116] The meanings of the parameters in Table 2 are as follows:

[0117] are the focal lengths of the left camera in the horizontal and vertical directions, respectively;

[0118] is the image principal point coordinate of the left camera;

[0119] is the radial distortion coefficient of the left camera;

[0120] is the tangential distortion coefficient of the left camera;

[0121] are the focal lengths of the right camera in the horizontal and vertical directions, respectively;

[0122] is the image principal point coordinate of the right camera;

[0123] is the radial distortion coefficient of the right camera;

[0124] is the tangential distortion coefficient of the right camera;

[0125] (T x , T y , T z ) are the translation vectors.

[0126] The average re-projection error of the three groups of images is shown in Table 3. The average re-projection error is calculated according to the calibration results.

[0127] Table 3 Average re-projection error of three groups of experiments

[0128]

[0129] According to Table 3, the average projection error of the third group of experiments shows that multi-stage preprocessing significantly optimizes the calibration accuracy. The untreated group has difficulty in feature extraction due to the problems of reflection interference, low contrast and corner blur in the original image, and the average re-projection error is 0.30 pixels. The traditional Harris enhancement group reduces the error to 0.25 pixels through bilateral filtering and CLAHE algorithm, verifying the effectiveness of the preprocessing process. The error of the Sobel-Harris fusion enhancement group is further reduced to 0.16 pixels, which is reduced by 36% compared with the traditional group and by 46.67% compared with the untreated group, fully proving the effectiveness of this method.

[0130] Compared with the prior art, the beneficial effects of the multi-stage feature enhancement-based long-distance binocular camera calibration optimization method of the embodiments of the present disclosure can be summarized as follows: the present disclosure aims to construct a multi-stage feature enhancement processing framework including reflection suppression, contrast enhancement and corner enhancement, and a Sobel-Harris fusion enhancement corner enhancement method, which improves the calibration results. In each embodiment, the multi-stage feature enhancement processing framework including reflection suppression, contrast enhancement and corner enhancement is constructed to solve the precision decay problem commonly existing in outdoor long-distance binocular camera calibration, and the traditional chessboard calibration method under complex lighting conditions is systematically solved. The proposed Sobel-Harris fusion enhancement corner intensity method effectively improves the robustness of feature extraction of the calibration board while maintaining the integrity of the image geometric structure, and significantly enhances the clarity of the long-distance chessboard corner.

[0131] Based on the above inventive concept, the multi-level feature enhancement based long-distance binocular camera calibration optimization method of various embodiments of the present disclosure at least processes the high light area existing on the calibration board, retains the image edge features while suppressing the mirror reflection noise, balances the smoothing processing and edge information retention; adopts the CLAHE algorithm combined with the block processing and the contrast gain threshold to enhance the checkerboard texture; adopts the bilateral filter to suppress the high light mirror reflection, strengthens the image corner points, optimizes the image gray scale distribution combined with the CLAHE algorithm to enhance the local contrast feature of the checkerboard texture, thereby proposing the multi-level feature enhancement based long-distance binocular camera calibration optimization method, suppressing the high light reflection and retaining the edge sharpness through the bilateral filtering, enhancing the texture saliency in the low light area combined with the CLAHE algorithm, and innovatively fusing the Sobel gradient direction and the Harris corner point response function to strengthen the corner point area detection capability, effectively solving the mirror reflection noise, low contrast and corner point blur problems, and improving the calibration precision.

[0132] The present disclosure also provides a multi-level feature enhancement based long-distance binocular camera calibration optimization device, which comprises one or more processing modules configured to execute the multi-level feature enhancement based long-distance binocular camera calibration optimization method described above, and is at least configured to execute the specific embodiments of steps S1 to S3.

[0133] The present disclosure also provides a computer readable storage medium having computer executable instructions stored thereon, wherein the computer executable instructions are executed by a processor to mainly implement the multi-level feature enhancement based long-distance binocular camera calibration optimization method described above, and at least comprise:

[0134] The high light area existing on the calibration board is processed to retain the image edge features while suppressing the mirror reflection noise, and to balance the smoothing processing and edge information retention.

[0135] The CLAHE algorithm combined with the block processing and the contrast gain threshold is adopted to enhance the checkerboard texture.

[0136] The bilateral filter is adopted to suppress the high light mirror reflection, the image corner points are strengthened, and the CLAHE algorithm is combined to optimize the image gray scale distribution to enhance the local contrast feature of the checkerboard texture.

[0137] The above embodiments are only exemplary embodiments of the present disclosure and are not used to limit the present disclosure, and the protection scope of the present disclosure is defined by the claims. Those skilled in the art can make various modifications or equivalent replacements to the present disclosure within the spirit and protection scope of the present disclosure, and such modifications or equivalent replacements are also regarded as falling within the protection scope of the present disclosure.

Claims

1. A calibration and optimization method for long-range binocular cameras based on multi-level feature enhancement, including: The calibration board has a highlight area that is processed to suppress specular reflection noise while preserving image edge features, thus balancing smoothing processing and edge information preservation. The checkerboard texture is enhanced by using the CLAHE algorithm combined with block processing and contrast gain thresholding. A bilateral filter is used to suppress specular reflections and enhance the corner points of the image. The CLAHE algorithm is combined to optimize the grayscale distribution of the image to enhance the local contrast features of the checkerboard texture.

2. The method according to claim 1, wherein, The calibration board has highlighted areas that need to be processed, including: The image is processed using a bilateral filter, and the weights of the bilateral filter are used to construct a fusion of spatial and brightness dimensions. By multiplying spatial and luminance weights, bilateral filtering preserves edge pixels based on the abrupt changes in luminance in the edge region while suppressing specular reflection noise.

3. The method according to claim 2, wherein: Spatial weight: through the formula Measure pixel position correlation, where σ s The influence of spatial distance on weights is adjusted. ||xy|| represents the Euclidean distance between two points at coordinates x and y, which measures the spatial interval between two pixels and ensures the spatial correlation constraint of neighboring pixels. Brightness weighting: using the formula Depicting pixel grayscale differences, σ r The effect of brightness changes on the weights is controlled, where I(x) is the pixel brightness value in the horizontal direction and I(y) is the pixel brightness value in the vertical direction. This makes the algorithm sensitive to abnormal brightness changes caused by reflections and specifically suppresses specular reflection interference.

4. The method according to claim 3, wherein, The filtering formula is: Ω is a neighborhood window centered at x; W total =∑ y∈Ω W s (x,y)·W r (x,y) represents the total weight.

5. The method according to claim 4, wherein, The checkerboard texture is enhanced using the CLAHE algorithm combined with block processing and contrast gain thresholding, including: The image is divided into grid sub-regions of a preset size, with the geometric center of each sub-region serving as the control point; Independently process the spatial coordinates and head / tail data allocation of sub-regions; By truncating the histogram through dynamic range limitation, the pixel intensity difference of the truncated part is evenly redistributed to the sub-region. This process may exceed the threshold again and cause secondary overflow. Through subsequent iterative optimization, the overflow amount is reduced to a preset tolerance threshold to ensure that the histogram shape is stable and converged. Ultimately, adaptive enhancement of local contrast in the checkerboard texture is achieved, and the output result is obtained.

6. The method according to claim 5, wherein, Also includes: The method of fusing the Sobel operator and the Harris corner response function enhances the response value of the corner region by extracting the gradient direction distribution features of the pixel neighborhood and fusing them with the Harris corner response function, thereby achieving robust feature enhancement.

7. The method according to claim 6, wherein, Sobel-Harris fusion enhancement corner point method includes: Calculate the gradient to obtain the gradient magnitude; Construct a local structure tensor and perform scale-space integration using Gaussian smoothing; Harris corner response calculation; Feature fusion and contrast enhancement utilize gradient magnitude features, processed through normalized linear fusion.

8. The method according to claim 7, wherein, Gradient calculation yields the gradient magnitude, including: Perform Sobel convolution on the input grayscale image I(x,y) to calculate the horizontal gradient field G. x With the vertical gradient field G y ; Based on the horizontal gradient field G x With the vertical gradient field G y The gradient magnitude is obtained.

9. The method according to claim 8, wherein, Construct the local structure tensor and perform scale-space integration using Gaussian smoothing, including: Matrix components and G xy =G x ⊙G y , where ⊙ represents element-wise multiplication; Gaussian window size by size is the Gaussian window size, and σ is the standard deviation of the Gaussian function.

10. The method according to claim 9, wherein, Feature fusion and contrast enhancement utilize gradient magnitude features, processed through normalized linear fusion, including: By introducing a coefficient β, a value that yields the best effect is determined. In the high-value area, local brightness attenuation is generated to enhance edge contrast; in the low-value area, the original gray-scale distribution is maintained to suppress noise amplification.

Images