A Fast On-site Calibration Method for Large Field-of-view Cameras Applicable to Complex Working Conditions

Abstract

The present invention discloses a method for on-site rapid calibration of a large field-of-view camera applicable to complex working conditions, including: on-site measurement scene arrangement, camera placement and adjustment, left camera acquisition of cross calibration target images, decoding of cross calibration target images, orientation of cross calibration target images, optimization of camera internal parameters, calibration of right camera internal parameters, synchronous acquisition of a frame of measurement scene images by the left and right cameras, feature point detection, feature point matching, solution of initial external parameter values, total station measurement of point distances, correction and optimization of camera external parameters, and export of calibration data. The present invention combines close-range photogrammetry technology, binocular stereo vision technology and total station ranging technology, uses a small-sized cross calibration target and the distance between a pair of feature points that are far apart in the measurement scene obtained by the total station as a flexible calibration target, realizes the calibration of a large field-of-view camera under complex working conditions, and can complete the on-site rapid calibration of the internal and external parameters of a large field-of-view camera with a field of view of dozens of meters to hundreds of meters within a few minutes.

Description

Technical Field

[0001] The present invention belongs to the technical field of 3D vision, and relates to a camera calibration method, in particular to a method for quickly calibrating a large field of view camera on-site under complex working conditions. Background Technique

[0002] The large field of view 3D vision detection technology is one of the effective means to realize the measurement of the size and deformation of large workpieces in the fields of aerospace, bridge tunnels, wind turbine blades, ships, etc. A high-precision and flexible large field of view camera calibration method is the cornerstone for realizing accurate and rapid measurement. Traditional camera calibration methods, such as Zhang Zhengyou's planar calibration method, Tsai's two-step method, etc., can obtain high-precision calibration results, but usually require the calibration target to occupy one-third or more of the measurement field of view. However, large-size calibration targets have many problems such as difficult manufacturing, high cost, inconvenient transportation, maintenance and use, making it difficult to implement large field of view camera calibration under complex on-site working conditions.

[0003] In the journal paper "Global Calibration of Large Field of View Multi-Camera Video Measurement System" (Optics and Precision Engineering, 2012, 20(02): 369-378), a large cross-shaped calibration target with concentric circular ring-shaped coding fiducial points was designed for the calibration of large field of view multi-cameras. This target is large in size, cumbersome to disassemble and assemble, difficult to carry and easily deformed, and increases proportionally with the increase of the field of view, and does not fundamentally solve the dependence of the large field of view on the size of the calibration object. Chinese invention patent ZL201510616889.0 arranges four sets of collinear constraint calibration rulers in a large visual measurement field of view, uses the invariant property of cross-ratio and linear constraints to solve the distortion parameters, linearly solves the initial values of the calibration parameters through the calibration control points in space, and finally combines the distortion coefficients and the calibration initial values to use the Levenberg-Marquardt (LM) optimization method to perform overall optimization with the minimum reprojection error as the objective function to obtain the accurate result of large field of view camera calibration. However, this method needs to rely on a third-party industrial measurement software to obtain the three-dimensional coordinates of the control points on the object to be measured, and the calibration accuracy depends on the third-party industrial measurement software, lacking independence.

[0004] Chinese invention patent ZL201810397640.9 measures the characteristic points of a marker by total station to obtain the three-dimensional coordinates of the characteristic points in the world coordinate system, uses the corner extraction and sub-pixel positioning methods to obtain the pixel positions of the characteristic points in the two-dimensional image coordinate system, and transforms the three-dimensional coordinates of the characteristic points in the world coordinate system to the camera coordinate system through rotation and translation; secondly, according to the projection, rotation, and translation transformation matrices, the mapping relationship between the pixel positions of the characteristic points in the two-dimensional image coordinate system and the three-dimensional coordinates in the camera coordinate system is established, and a non-linear error equation is established; finally, the internal parameters of the single camera and the external parameter matrix of the camera are obtained by iterative optimization. In this method, the three-dimensional coordinates rely entirely on the total station measurement, which is prone to introducing measurement errors. Moreover, as the scene increases and the number of markers increases, the measurement steps will become increasingly cumbersome; secondly, the mapping relationship between the three-dimensional points and the two-dimensional image coordinates is established only through the rotation and translation transformation matrices and the projection matrix. If the lens distortion is large, it is easy to cause errors in the mapping relationship and the robustness is insufficient.

[0005] The journal paper "A Large Field-of-View Stereo Vision Calibration Method Based on Defocus Images" (Journal of Applied Sciences, 2019, 37(06): 795-805) aims at the on-site calibration problem of helicopter blade motion parameter measurement and proposes a large field-of-view stereo camera calibration method based on defocus images. This method first takes multiple groups of targets located at the defocus positions of the small field of view, and combines the Levenberg-Marquardt (LM) algorithm to complete the solution of the internal parameters. Then, the calibration target is placed multiple times at the large field of view measurement position to make the target image fill the entire calibration field of view as much as possible, and the external parameters of the camera are iteratively optimized by combining the LM algorithm. In the external parameter calibration link of this method, it is necessary to hold the calibration target at different postures and place it multiple times at the large field of view measurement position. In the paper, a lift is used to implement this process, which is not only cumbersome to operate, but also increases potential safety hazards to a certain extent. Summary of the Invention

[0006] The present invention combines close-range photogrammetry technology, binocular stereo vision technology and total station ranging technology, and proposes a method for quickly calibrating a large field-of-view camera on-site applicable to complex working conditions. This method uses a small-sized cross calibration target and the distance between a pair of relatively far-apart landmark points in the measurement scene obtained by the total station as a flexible calibration target, overcoming the restriction of the calibration target size on the calibration of large field-of-view cameras: First, using the small-sized cross calibration target, combined with binocular stereo vision technology and close-range photogrammetry technology, quickly and accurately calibrate the internal parameters of the camera; Second, based on the obtained internal parameters of the camera and the distance between a pair of landmark points measured by the total station, based on close-range photogrammetry technology, accurately calibrate the external parameters of the large field-of-view camera. Since there is no need to manufacture a large calibration target, the method of the present invention is not restricted by the calibration target size and can calibrate large field-of-view cameras with a range of dozens of meters to hundreds of meters. In short, compared with the existing ones, the method of the present invention fundamentally solves the restriction of the calibration target size on the calibration of large field-of-view cameras, with a simple calibration process, fast calibration speed and strong universality.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] A method for quickly calibrating a large field-of-view camera on-site applicable to complex working conditions, comprising the following steps:

[0009] The first step, on-site measurement scene layout

[0010] Paste circular non-coded landmark points at key positions on the surface of the object to be measured in the measurement scene according to the actual measurement requirements;

[0011] The second step, camera placement and adjustment

[0012] According to the arranged measurement scene, determine the placement distance and angle between the two cameras, and adjust the focal length so that the cameras are focused at infinity;

[0013] The third step, the left camera captures images of the cross calibration target

[0014] Control the left camera to capture images of the cross calibration target at different positions and postures;

[0015] The fourth step, decoding of the cross calibration target images

[0016] Extract the coordinates of the landmark points in the cross calibration target images captured in the third step, and then perform decoding to obtain the landmark point IDs;

[0017] The fifth step, orientation of the cross calibration target images

[0018] Select two decoded cross calibration target images in the fourth step for relative orientation; after successful orientation, perform absolute orientation on the other images decoded in the fourth step, and then obtain the exterior orientation elements and the three-dimensional coordinates of the landmark points of all images;

[0019] Step 6, Optimization of Camera Internal Parameters

[0020] Taking the reprojection error as the objective equation, the internal parameters of the camera, the external orientation elements of the cross calibration target image obtained in the fifth step, and the three-dimensional coordinates of the fiducial points are used as the parameters to be optimized for overall bundle adjustment to obtain the optimal internal parameters of the camera;

[0021] Step 7, Calibration of Right Camera Internal Parameters

[0022] Calibrate the internal parameters of the right camera according to the methods in the third to sixth steps;

[0023] Step 8, Synchronously Acquire One Frame of Measurement Scene Image by Left and Right Cameras

[0024] Control the left and right cameras to synchronously acquire one frame of the measurement scene image arranged in the first step;

[0025] Step 9, Fiducial Point Detection

[0026] Detect the center coordinates of the circular non-coded fiducial points in the measurement scene images acquired by the left and right cameras;

[0027] Step 10, Fiducial Point Matching

[0028] Adopt the shape context algorithm to achieve fast matching of the circular non-coded fiducial points detected in the ninth step;

[0029] Step 11, Solution of Initial Values of External Parameters

[0030] Taking the left camera coordinate system as the reference coordinate system, that is, the rotation matrix in the external parameters of the left camera is the identity matrix and the elements of the translation matrix are all 0. According to the fiducial point matching results in the tenth step, use the relative orientation algorithm to obtain the initial values of the external parameters of the right camera; then, reconstruct the three-dimensional coordinates of the circular non-coded fiducial points in the measurement scene;

[0031] Step 12, Total Station Measurement of Point Distance

[0032] Use a total station to measure the distance between two relatively far circular non-coded fiducial points in the measurement scene reconstructed in the eleventh step, where the distance between the two measured circular non-coded fiducial points should be more than half of the size of the measurement scene in the length direction;

[0033] Step 13, Calibration and Optimization of Camera External Parameters

[0034] Taking the point distance measured in the twelfth step as the scale, calculate the scaling factor, calibrate the external parameters of the right camera solved in the eleventh step, and optimize the external parameters of the right camera using the bundle adjustment method;

[0035] Step 14, Export Calibration Data

[0036] Export the internal and external parameters of the left and right cameras to complete camera calibration.

[0037] A specific embodiment, in the second step, the method for placing and adjusting the cameras is as follows: the included angle θ between the two cameras is in the range of 20° to 30°, that is, 20° < θ < 30°, the measured scene size is V x ×V y , in units of mm; the camera resolution is W×H; the pixel size is px, in units of μm / pixel; the focal length of the camera lens is f, in units of mm; the calculation formula for the distance D from the two cameras to the object to be measured is: D≈Vx×f / (W×px), in units of mm; the calculation formula for the placement distance L between the two cameras is: L≈2×D×sin(θ / 2); after the cameras are placed, adjust the focal length, that is, rotate the focus ring to the infinity mark [∞] on the lens, so that both cameras can clearly image within a certain distance from the hyperfocal point to the measured object.

[0038] A specific embodiment, in the third step, the cross calibration target is a cross structure with four equal-length arms composed of profile I and profile II that are equal in length and perpendicular to each other; N mark points, N≥4, are printed or pasted at equal intervals on the center line position of one of the arms belonging to profile I; N + 1 mark points are printed or pasted at equal intervals on the center line positions of the other three arms; the cross ratio of every 4 adjacent mark points on each arm of the target is equal to 4 / 3; the mark points printed or pasted on each profile of the target are all circular non-coded mark points of the same size; the mark points printed or pasted on each profile of the target are all collinear.

[0039] A specific embodiment, the coding rule of the cross calibration target is as follows: for profile I with 2N + 1 mark points printed or pasted, the IDs of the mark points are sequentially assigned from 0 to 2N in the direction from the starting end to the terminal end; for profile II with 2N + 2 mark points printed or pasted, the IDs of the mark points are sequentially assigned from 2N + 1 to 4N + 2 in the direction from the starting end to the terminal end; the starting end and the terminal end of profile I and profile II are defined as follows: the starting end of profile I is defined as the end where N mark points are printed or pasted on the arm, and the terminal end is defined as the end where N + 1 mark points are printed or pasted on the arm; rotate profile I counterclockwise by 90° to coincide with profile II, then the end of profile II that coincides with the starting end of profile I is defined as the starting end of profile II, and the end that coincides with the terminal end of profile I is defined as the terminal end of profile II.

[0040] A specific embodiment, in the fourth step, the cross calibration target image decoding process is as follows:

[0041] 1) Detection of circular non-coded mark points

[0042] Extract the center coordinates of the mark points in the cross calibration target image collected in the third step;

[0043] 2) Straight line fitting

[0044] Use the RANSAC algorithm to successively fit two straight lines to the extracted central coordinates and mark the inliers corresponding to each straight line;

[0045] 3) Straight line segmentation

[0046] Calculate the intersection point of the two fitted straight lines and use it as the segmentation point to divide each straight line into two segments; Divide the inliers of each marked straight line into two groups, corresponding to the two segments of the straight line after segmentation respectively;

[0047] 4) Cross-ratio invariance test

[0048] Calculate the cross-ratio of the inliers of each segmented straight line and mark the inliers that meet the cross-ratio invariance condition as the landmark points on the segmented straight line;

[0049] 5) Verification of the number of landmark points

[0050] Statistically check whether the number of landmark points on the four segmented straight lines after the cross-ratio invariance test is consistent with the number of landmark points printed or pasted on the four arms of the cross calibration target described in the third step: If the number of landmark points on one of the segmented straight lines is N and the number of landmark points on the other three segmented straight lines is N + 1, it is determined to be qualified;

[0051] 6) Assignment of landmark point IDs

[0052] If the verification of the number of landmark points is qualified, assign IDs to the landmark points on the four segmented straight lines in sequence.

[0053] In a specific embodiment, the detection of circular non-coded landmark points is implemented by the following real-time detection algorithm for circular non-coded landmark points:

[0054] First, perform initialization on the CPU side and the GPU side, that is, allocate enough pinned memory on the CPU side to store the target image collected in the third step, and allocate enough memory area in the global memory of the GPU side for storing data during the GPU side processing;

[0055] Secondly, upload the target image collected in the third step from the CPU side to the GPU side. Asynchronous transmission is used during transmission. After the image transmission is completed, start binaryzation processing on the GPU side. After binaryzation is completed, use a two-step method to perform connected component labeling and analysis on the GPU side, and extract the integer pixel edges of circular non-coded landmark points on this basis;

[0056] Next, extract sub-pixel edges on the GPU side based on the spatial moment method, and use the least squares ellipse fitting algorithm to obtain the central coordinates of circular non-coded landmark points;

[0057] Finally, the detection results are asynchronously transferred from the GPU side to the CPU side to complete the real-time detection of the fiducial points in a calibration target image collected in the third step.

[0058] A specific embodiment, the method for cross-ratio invariance test is as follows: for the inliers of a segmented line, if the number is less than N, it is determined that all inliers do not meet the cross-ratio invariance condition; if the number is equal to or more than N, calculate the distance from each inlier to the intersection point to be found, and sort them in ascending order of distance; take the first 4 inliers after sorting, calculate the cross-ratio λ, if it meets the cross-ratio invariance condition: 4 / 3 - ε ≤ λ ≤ 4 / 3 + ε, it is determined that the first 4 inliers all meet the invariance condition, and mark the first 4 inliers as the fiducial points on the segmented line; for the 5th inlier after sorting, take the first three inliers adjacent to it, that is, the 4th, 3rd, and 2nd inliers after sorting, to form a set of adjacent 4 inliers, calculate the cross-ratio λ, if it meets the cross-ratio invariance condition: 4 / 3 - ε ≤ λ ≤ 4 / 3 + ε, it is determined that the 5th inlier after sorting meets the invariance condition, and mark it as the fiducial point on the segmented line; finally, the other inliers after sorting are sequentially tested according to the determination method of the 5th inlier; where ε represents the error factor, and the value range of ε is: ε = 0.02 - 0.08.

[0059] A specific embodiment, the method for fiducial point ID assignment is as follows: for a segmented line with N fiducial points, calculate the distance from each fiducial point on it to the intersection point to be found, and sort the fiducial points in descending order of distance, and then assign the IDs of the fiducial points as 0 to N - 1 in sequence after sorting; for another segmented line collinear with a segmented line containing N fiducial points, calculate the distance from each fiducial point on it to the intersection point to be found, and sort the fiducial points in ascending order of distance, and then assign the IDs of the fiducial points as N to 2N in sequence after sorting; secondly, construct a vector with the starting point being the fiducial point with ID 0 and the ending point being the fiducial point with ID 2N, and use this vector as the reference vector; for the other two segmented lines, calculate the distance between the fiducial points on each line and the intersection point to be found respectively, and select the fiducial point with the farthest distance on each line, connect it with the fiducial point with ID 0 to form a vector, where the fiducial point with ID 0 is the starting point of the vector, and then take the cross product with the constructed reference vector. If the cross product is positive, sort all the fiducial points on this segmented line in descending order of the calculated distance, and assign the IDs of the fiducial points as 2N + 1 to 3N + 1 in sequence after sorting; if the cross product is negative, sort all the fiducial points on this segmented line in ascending order of the calculated distance, and assign the IDs of the fiducial points as 3N + 2 to 4N + 2 in sequence after sorting.

[0060] A specific embodiment. In the fifth step, when performing image orientation, the included angle ρ between two images along the optical axis direction is calculated according to the rotation matrix. If the included angle ρ > 20°, the relative orientation is successful; otherwise, the orientation fails, and another two images are used for relative orientation until it is successful. After the relative orientation is successful, the three-dimensional coordinates of the fiducial points in the two images are calculated, and these fiducial points are used as control points. The pyramid method is used to perform absolute orientation on other images to obtain the exterior orientation elements of other images.

[0061] A specific embodiment. In the thirteenth step, the method for camera external parameter calibration and optimization is as follows: The point distance measured by the total station in the twelfth step is used as the scale, the scaling factor is calculated, and the three-dimensional coordinates of all fiducial points in the measurement scene and the initial values of the external parameters of the right camera solved in the eleventh step are scaled to obtain the calibrated external parameters of the right camera. Finally, the bundle adjustment method is used, with the reprojection error as the optimization target, and the optimal camera external parameters are iteratively solved.

[0062] Compared with the prior art, the beneficial effects of the present invention are:

[0063] (1) The present invention uses a small-sized cross calibration target and the distance between a pair of fiducial points far apart in the measurement scene obtained by the total station as a flexible calibration target, solving the bottleneck problems such as the difficulty in manufacturing large-sized calibration targets, high costs, inconvenient transportation, maintenance, and use in large field-of-view camera calibration.

[0064] (2) Since there is no need to manufacture large calibration targets, the method of the present invention is not restricted by the size of the calibration target and can achieve the calibration of large field-of-view cameras with a field of view of dozens of meters to over a hundred meters.

[0065] (3) The calibration process of the method of the present invention is simple and the calibration speed is fast, and it can achieve on-site rapid calibration of large field-of-view cameras within a few minutes under complex working conditions.

[0066] (4) The present invention combines the close-range photogrammetry technology and the binocular stereo vision technology, with high calibration accuracy. The calibration accuracies of the internal and external parameters of large field-of-view cameras can both be controlled within 0.1 pixel (reprojection error). Description of the Drawings

[0067] Figure 1 It is a flow chart of the specific operation steps of the present invention.

[0068] Figure 2 It is a method device diagram of the present invention.

[0069] Figure 3 It is a circular non-coded fiducial point with a crosshair printed in the center.

[0070] Figure 4 It is an arrangement diagram of the measurement scene at the site of deformation detection of a certain large wind turbine blade.

[0071] Figure 5 It is a schematic diagram of the binocular camera arrangement.

[0072] Figure 6 It is a schematic diagram of a cross calibration target.

[0073] Figure 7 They are cross calibration target images collected by the left camera at different positions and with different postures.

[0074] Figure 8 It is a flowchart for decoding cross calibration target images.

[0075] Figure 9 It is a flowchart for real-time detection of circular non-coded fiducial points.

[0076] Figure 10 It is Figure 7 The decoding result of a cross calibration target image in

[0077] Figure 11 They are cross calibration target images collected by the right camera at different positions and with different postures.

[0078] Figure 12 It is a frame of Figure 4 The on-site measurement scene image arranged in, where a is the image collected by the left camera and b is the image collected by the right camera.

[0079] Figure 13 It is Figure 12 The detection and matching results of circular non-coded fiducial points in Figure 12 Among them, a is the detection and matching result of the image collected by the left camera a in Figure 12 And b is the detection and matching result of the image collected by the right camera b in Detailed implementation method

[0080] The following will describe in detail the implementation method of the present invention in combination with the accompanying drawings and embodiments.

[0081] The present invention proposes a method for rapid on-site calibration of a large field-of-view camera applicable to complex working conditions. The implementation process of this method is as Figure 1 shown, and the device used is as Figure 2 shown. The device mainly consists of a binocular CCD camera, a trigger, a high-performance computer, a total station, a cross calibration target, circular non-coded fiducial points, a tripod, a pan-tilt head, etc.

[0082] Referring to Figure 1 , the main process of the method for rapid on-site calibration of a large field-of-view camera applicable to complex working conditions according to the present invention is as follows:

[0083] The first step is to arrange the on-site measurement scenario. According to the actual measurement requirements, paste circular non-coded fiducial points at key positions on the surface of the object to be measured within the measurement scenario.

[0084] For example, Figure 3 as shown, a crosshair for cooperating with the total station to measure distance is printed at the center of the circular non-coded fiducial point used. In addition, the minimum radius of the circular non-coded fiducial point imaged on the image needs to be greater than 3 pixels. As Figure 4 shown, some circular non-coded fiducial points (black background with white dots) with crosshairs are arranged on the wall of a certain factory building. The inner diameter of the fiducial point (radius of the white dot) is 200 mm, and the radius imaged on the image is approximately 7 - 8 pixels. These fiducial points are used for camera calibration and coordinate transformation (transforming the measurement coordinate system to the factory building coordinate system). Fiducial points with a radius of 300 mm are pasted on the surface of the blade to be measured, and the radius imaged on the image is approximately 12 - 13 pixels. These fiducial points are the points to be measured and are also used for camera calibration.

[0085] The second step is to place and adjust the cameras. According to the measurement scenario arranged in the first step, determine the placement distance and angle between the two cameras. As Figure 5 shown, the included angle θ between the two cameras is generally within the range of 20° to 30° (20° < θ < 30°). Assuming the size of the measurement scenario is V x ×V y (mm), the camera resolution is W × H, the pixel size is px (μm / pixel), and the focal length of the camera lens is f (mm), then the calculation formula for the distance D (mm) from the two cameras to the object to be measured is: D ≈ Vx × f / (W × px); the calculation formula for the placement distance L between the two cameras is: L ≈ 2 × D × sin(θ / 2). After the cameras are placed, adjust the focal length, that is, rotate the focus ring to the infinity mark [∞] on the lens, so that both cameras can clearly image within a certain distance from the hyperfocal point to the object to be measured. Figure 4 In the measurement scenario in, the size is approximately 42 meters × 30 meters, that is, the Vx size is at least 42 meters. The two Basler acA2440 - 75um area array cameras used have a resolution of: 2448 pixel × 2048 pixel, the pixel size is 3.45 μm / pixel, and the lens is a Computar 8 mm fixed-focus lens. When the camera included angle is 25°, it can be calculated that the distance from the two cameras to the blade to be measured is approximately 40 m, and the distance between the two cameras is approximately 17 m.

[0086] The third step is for the left camera to collect the cross calibration target image. As Figure 6As shown in the figure, the cross calibration target used is a cross shape with four equal-length arms formed by two profiles of equal length that intersect perpendicularly. Profile I and Profile II (made of aluminum alloy or carbon fiber, etc.) are connected by bolts. On the center line of one of the arms belonging to Profile I, N marking points are printed or pasted at equal intervals, where N≥4. On the center lines of the other three arms, N + 1 marking points are printed or pasted at equal intervals. The cross ratio of every four adjacent marking points on each arm of the target is equal to 4 / 3. The marking points printed or pasted on each profile of the target are all circular non-coded marking points of the same size. The marking points printed or pasted on each profile of the target need to be collinear. Figure 6 In this case, N is 6, and there are a total of 27 marking points on the target.

[0087] The ID coding rule for the marking points on the cross calibration target adopted is as follows: For Profile I on which 2N + 1 marking points are printed or pasted, starting from the starting end to the terminal end, the IDs of the marking points are sequentially assigned from 0 to 2N. For Profile II on which 2N + 2 marking points are printed or pasted, starting from the starting end to the terminal end, the IDs of the marking points are sequentially assigned from 2N + 1 to 4N + 2. The starting end and the terminal end of Profile I and Profile II are defined as follows: The starting end of Profile I is defined as the end on which N marking points are printed or pasted on the arm, and the terminal end is defined as the end on which N + 1 marking points are printed or pasted on the arm. Assuming that Profile I is rotated 90° counterclockwise to coincide with Profile II, then the end of Profile II that coincides with the starting end of Profile I is defined as the starting end of Profile II, and the end that coincides with the terminal end of Profile I is defined as the terminal end of Profile II.

[0088] Control the left camera to collect cross calibration target images in different positions and postures, such as the target facing the camera directly, pulling the target closer to the camera by a certain distance and facing the camera directly, pulling the target farther from the camera by a certain distance and facing the camera directly, tilting the target forward by 30°, tilting the target backward by 30°, rotating the target clockwise by 180°, moving the target upward by a certain distance and facing the camera directly, moving the target downward by a certain distance and facing the camera directly, etc. It should be noted that the cross calibration target needs to occupy about 1 / 3 of the camera's field of view. Figure 7 As shown in Figure 4 the cross calibration target image collected by the left camera arranged in the measurement scene shown in

[0089] In the fourth step, decode the cross calibration target image. First, extract the coordinates of the marking points in the cross calibration target image collected in the third step. Secondly, perform decoding to obtain the ID of the marking point. The specific implementation process is as Figure 8 shown, including the following steps:

[0090] 1) Adopt methods such as the real-time detection algorithm for circular non-coded marking points to extract the center coordinates of the marking points in the target image collected in the third step. The flow of the real-time detection algorithm for circular non-coded marking points is as Figure 9As shown, it is mainly implemented on the GPU side. First, initialize on both the CPU side and the GPU side, that is, allocate enough pinned memory on the CPU side to store the target images collected in the third step, and allocate enough memory areas in the global memory on the GPU side for data storage during the GPU-side processing. The initialization on both the CPU side and the GPU side takes a relatively long time, but it only needs to be initialized once, that is, this step will not be executed again after the first initialization is completed. Secondly, upload the target images collected in the third step from the CPU side to the GPU side, and use asynchronous transmission during the transmission. After the image transmission is completed, start the binarization process on the GPU side. The binarization on the GPU side uses the Sauvola local adaptive binarization algorithm. After the binarization is completed, use the two-pass method (Two-Pass algorithm) on the GPU side to perform connected component labeling and analysis, and based on this, extract the integer-pixel edges of the circular non-coded fiducial points; next, on the GPU side, extract the sub-pixel edges based on the spatial moment method, and use the least squares ellipse fitting algorithm to obtain the center coordinates of the circular non-coded fiducial points; finally, asynchronously transmit the detection results from the GPU side to the CPU side to complete the real-time detection of the fiducial points in a calibration target image collected in the third step.

[0091] 2) Line fitting. Use the RANSAC (Random Sample Consensus) algorithm to successively fit two lines to the coordinates of the circular non-coded fiducial points detected in step 1), and mark the inliers corresponding to each line. The inlier marking method is as follows: traverse all the fiducial points obtained in step 1), and calculate the distance D from each fiducial point to each of the fitted lines; if the distance from the fiducial point to a certain fitted line is less than the threshold T, that is, D < T, then mark this fiducial point as an inlier of the corresponding line. After the inlier marking is completed, count the number of inliers for each line. The threshold T is determined by the user, and generally takes a value of 2 to 5 pixel sizes.

[0092] 3) Line segmentation. Calculate the intersection point of the two lines fitted in step 2), and use it as the segmentation point to divide each line into two segments. In addition, divide the inliers of each line marked in step 2) into two groups, corresponding to the two segments of the line after segmentation. In fact, the four segments of the line after segmentation correspond one by one to the center lines of the four arms of the cross calibration target.

[0093] 4) Cross-ratio invariance test. Calculate the cross-ratio of the inliers of each straight line segment segmented in step 3), and mark the inliers that meet the cross-ratio invariance condition as the landmark points on the segmented straight line. The cross-ratio invariance test method is as follows: For the inliers of a segmented straight line, if the number is less than N, it is determined that all inliers do not meet the cross-ratio invariance condition; if the number is equal to or more than N, calculate the distance from each inlier to the intersection point to be found, and sort them in ascending order of distance; take the first 4 inliers after sorting, calculate the cross-ratio λ, if it meets the cross-ratio invariance condition: 4 / 3 - ε ≤ λ ≤ 4 / 3 + ε, it is determined that the first 4 inliers all meet the invariance condition, and mark the first 4 inliers as the landmark points on the segmented straight line; for the 5th inlier after sorting, take the three inliers adjacent to it, that is, the 4th, 3rd, and 2nd inliers after sorting, to form a set of 4 adjacent inliers, calculate the cross-ratio λ, if it meets the cross-ratio invariance condition: 4 / 3 - ε ≤ λ ≤ 4 / 3 + ε, it is determined that the 5th inlier after sorting meets the invariance condition, and mark it as the landmark point on the segmented straight line; finally, check the other inliers after sorting in turn according to the determination method of the 5th inlier. Among them, N represents the number of landmark points printed or pasted on the arm on one side of the starting end of profile I in step 3), ε represents the error factor, and the value range of ε is: ε = 0.02 - 0.08.

[0094] 5) Verification of the number of landmark points. Statistically verify whether the number of landmark points on the four segmented straight lines after the cross-ratio invariance test in step 4) is consistent with the number of landmark points printed or pasted on the four arms of the cross calibration target in step 3); if the number of landmark points on one of the segmented straight lines is N, and the number of landmark points on the other three segmented straight lines is N + 1, it is determined to be qualified. Among them, N represents the number of landmark points printed or pasted on the arm on one side of the starting end of profile I in step 3).

[0095] 6) Marker point ID assignment. If the number of marker points in step 5) is verified to be qualified, the IDs of the marker points on the straight lines after the four-segment division are assigned in sequence. The method for assigning marker point IDs is as follows: For a straight line after division with N marker points, calculate the distance from each marker point on it to the intersection point obtained in step 3), and sort the marker points in descending order of the distance. After sorting, assign the IDs of the marker points as 0 to N-1 in sequence; for another straight line after division that is collinear with a straight line after division containing N marker points, calculate the distance from each marker point on it to the intersection point obtained in step 3), and sort the marker points in ascending order of the distance. After sorting, assign the IDs of the marker points as N to 2N in sequence; secondly, construct a vector with the starting point being the marker point with ID 0 and the ending point being the marker point with ID 2N, and use this vector as the reference vector; for the other two straight lines after division, calculate the distance from each marker point on each straight line to the intersection point obtained in step 3), and select the marker point with the farthest distance on each straight line, connect it with the marker point with ID 0 to form a vector (the marker point with ID 0 is the starting point of the vector), and then take the cross product with the constructed reference vector. If the cross product is positive, sort all the marker points on this straight line after division in descending order of the calculated distance, and assign the IDs of the marker points as 2N+1 to 3N+1 in sequence after sorting; if the cross product is negative, sort all the marker points on this straight line after division in ascending order of the calculated distance, and assign the IDs of the marker points as 3N+2 to 4N+2 in sequence after sorting. Figure 10 shown as Figure 7 the decoding result of a cross calibration target image in

[0096] Step 5, image orientation. Select any two cross calibration target images decoded in step 4), and establish a mathematical model according to the coplanarity condition equation to complete the relative orientation between the images. During relative orientation, usually take the image space coordinate system of one of the images as the reference coordinate system (that is, the rotation matrix in the external orientation elements of this image is the identity matrix, and the elements of the translation matrix are all 0), and find the external orientation elements of the other image relative to the reference coordinate system. The mathematical model used is:

[0097]

[0098] In the formula, C is a non-linear function, [t x ,t y ,t z is the translation matrix between the two images, [x'2,y'2,z'2] T =R[x2,y2,z2] T , R is the rotation matrix between the two images, [x1,y1,z1] T , [x2,y2,z2] TThey are the coordinates of the fiducial points with the same ID in the two images in the image space coordinate system. The coordinates of the fiducial points in the image space coordinate system can be obtained by converting the coordinates of the fiducial points extracted in the fourth step through the coordinate system transformation in photogrammetry.

[0099] The rotation matrix and translation matrix between the two images can be obtained by solving the non-linear function C. Assume the rotation matrix R is

[0100]

[0101] where r1 to r9 represent the 9 elements of the rotation matrix. According to the rotation matrix, the angle ρ between the two images along the optical axis direction can be calculated as ρ = -arcsin(r6), and arcsin() represents the arcsine function. If the angle ρ > 20°, the relative orientation is successful; otherwise, the orientation fails, and another two images are used for relative orientation until it is successful. After the relative orientation is successful, the three-dimensional coordinates of the fiducial points in the two images are calculated, and these fiducial points are used as control points. The absolute orientation of other images is performed using the pyramid method to obtain the exterior orientation elements of other images.

[0102] Step 6, optimization of the camera internal parameters. Using the bundle adjustment algorithm, the camera internal parameters, the exterior orientation elements of the images, and the three-dimensional coordinates of the fiducial points are used as the parameters to be optimized, and the reprojection error is used as the objective equation. The accurate camera internal parameters and the three-dimensional coordinates of the fiducial points are obtained through gradual iteration. During the solution process, the initial camera distortion is set to zero and the initial values of the internal parameters are set to the factory settings of the camera. After the optimization of the camera internal parameters is completed, if the reprojection error ζ is less than the pre-given threshold τ, that is, ζ < τ, the calibration of the camera internal parameters is successful; otherwise, the calibration fails, and steps 3 to 5 are repeated until the calibration of the camera internal parameters is successful. The threshold τ generally takes a value of 0.5 to 1.0.

[0103] Thus, through steps 3 to 6, the calibration of the left camera internal parameters is completed.

[0104] Step 7, calibration of the right camera internal parameters. According to the left camera internal parameter calibration process described in steps 3 to 6, the internal parameters of the right camera are calibrated. It should be noted that the number of cross calibration target images collected by the right camera must be the same as the number of images collected by the left camera. Figure 11 Shown is the cross calibration target image collected by the right camera arranged in the Figure 4 measurement scene shown in, which is the same as the number of cross calibration target images collected by the left camera in Figure 7 .

[0105] Step 8: The left and right cameras synchronously capture a frame of the measurement scene image. If the internal parameters of both the left and right cameras are successfully calibrated, the camera trigger is used to trigger the left and right cameras to synchronously capture a frame of the measurement scene image arranged in Step 1. The measurement scene image captured by the left camera is called the left view, and the measurement scene image captured by the right camera is called the right view. Figure 12 As shown in a and b respectively in Figure 4 is a frame of the measurement scene image shown in

[0106] Step 9: Feature point detection. Extract the center coordinates of the circular non-coded feature points in the measurement scene images captured by the left and right cameras in Step 8. This can be achieved by methods such as the real-time detection algorithm for circular non-coded feature points described in Step 4.

[0107] Step 10: Feature point matching. The shape context algorithm is used to achieve fast matching of the circular non-coded feature points in the left and right views detected in Step 9. The specific implementation process is as follows:

[0108] 1) Let the set of feature points in the left view detected in Step 9 be P = {p1, p2..., p n}, where n represents the number of feature points. First, arbitrarily select one point p i as the reference point, 1 ≤ i ≤ n, and establish a polar coordinate system with p i as the pole. Secondly, create a circle with the pole as the center and radius R, where R = log(max(p j - p i ))), 1 ≤ j ≤ n, and max() represents the maximum value function. The circle is divided into k1×k2 sub-regions by equally dividing it into k1 parts in the radial direction according to the logarithmic distance and k2 parts in the angular direction. Among them, k1 is generally taken as 5 - 7, and k2 is generally taken as 10 - 12.

[0109] 2) Count the number of the remaining points in P in the sub-regions divided in Step 1) to obtain the histogram h i (k), 1 ≤ k ≤ k1×k2, which is called the shape context of point p i .

[0110] 3) Using the remaining feature points in P as reference points, repeat Steps 1) and 2) to calculate the shape contexts of the remaining feature points.

[0111] 4) Let the set of feature points in the right view be Q = {q1, q2..., q m}, where m represents the number of feature points. According to the process described in Steps 1) to 3), calculate the shape context of each feature point in the right view.

[0112] 5) Calculation of the similarity metric matrix. The value of the element in the i-th row and j-th column of the similarity metric matrix C is C ijThe calculation method is as follows:

[0113]

[0114] Among them, h i (k) and h j (k) represent the shape contexts of the fiducial point p i in the left view and the fiducial point q j in the right view respectively, where 1 ≤ i ≤ n and 1 ≤ j ≤ m.

[0115] 6) Based on the similarity metric matrix C calculated in step 5), the bipartite graph matching algorithm is used to solve the optimal matching π. The optimization objective function when solving the optimal matching π is:

[0116]

[0117] Among them, π represents the correspondence between the fiducial points in the left view and the fiducial points in the right view, and π(i) represents the i-th fiducial point p i in the left view corresponding to the π(i)-th fiducial point q π(i) in the right view, where 1 ≤ π(i) ≤ m. The π corresponding to the minimum value of the objective function H(π) is the optimal matching.

[0118] 7) Assign the ID of 100 + i to the pair of successfully matched fiducial points p i and q π(i) . Figure 13 As shown by a and b in Figure 12 are the fiducial point detection and matching results in the two images in

[0119] The eleventh step, solving the initial value of the external parameters. Taking the left camera coordinate system as the reference coordinate system (that is, the rotation matrix in the external parameters of the left camera is the identity matrix and the elements of the translation matrix are all 0), according to the fiducial point matching result in the tenth step, the initial value of the external parameters of the right camera is obtained by using the relative orientation method described in the fifth step. Then, according to the internal and external parameters of the left and right cameras, the three-dimensional coordinates of the circular non-coded fiducial points in the measurement scene are reconstructed.

[0120] The twelfth step, measuring the distance between points with a total station. Use a total station to measure the distance between two relatively far circular non-coded fiducial points in the measurement scene reconstructed in the eleventh step. The total station is generally placed between the two cameras, and the distance between the two measured circular non-coded fiducial points should be greater than half of the dimension in the length direction of the measurement scene. During measurement, the total station aims at the crosshair printed at the center of the circular non-coded fiducial point.

[0121] Step 13: Camera extrinsic parameter calibration and optimization. Using the point distances measured by the total station in Step 12 as the scale, calculate the scaling factor, scale the 3D coordinates of all the fiducial points in the measurement scene and the initial values of the extrinsic parameters of the right camera solved in Step 11 to obtain the calibrated extrinsic parameters of the right camera. Finally, using the bundle adjustment method, with the reprojection error as the optimization objective, iteratively solve for the optimal camera extrinsic parameters.

[0122] Step 14: Export the calibration data. Export the intrinsic and extrinsic parameters of the left and right cameras, and the camera calibration is completed. Taking the measurement scene shown in Figure 4 as an example, excluding the time for on-site measurement scene layout and camera placement and adjustment, the time consumed for the entire calibration process is within 8 minutes, and the camera calibration accuracy (reprojection error) is 0.073 pixel, indicating that the method of the present invention can meet the requirements of fast and high-precision on-site calibration of large-field cameras under complex working conditions. The finally exported internal and external parameters of the camera are shown in Table 1. In the table, f is the focal length, x0 and y0 are the principal point deviations in the x and y directions, k1, k2, and k3 are the radial distortion coefficients, b1 and b2 are the tangential distortion coefficients, p1 and p2 are the in-plane distortion coefficients. R is the rotation matrix in the camera extrinsic parameters, and T is the translation matrix in the camera extrinsic parameters.

[0123] Table 1 Calibration results of the internal and external parameters of the binocular camera arranged in the measurement scene shown in Figure 4

[0124]

[0125] In summary, the present invention combines the techniques of close-range photogrammetry, binocular stereo vision, and total station ranging technology, uses a small-sized cross calibration target and the distance between a pair of fiducial points far apart in the measurement scene obtained by the total station as a flexible calibration target, and realizes the calibration of large-field cameras under complex working conditions. The method of the present invention fundamentally solves the restriction of the calibration target size on the calibration of large-field cameras. The calibration process is simple, the calibration speed is fast, and the universality is strong. It can complete the on-site rapid calibration of the internal and external parameters of large-field cameras dozens of meters to hundreds of meters within a few minutes.​

Claims

1. A method for on-site rapid calibration of a large field-of-view camera applicable to complex working conditions, characterized in that, It includes the following steps: The first step is the on-site measurement scene layout Paste circular non-coded fiducial points at key positions on the surface of the object to be measured within the measurement scene according to the actual measurement requirements; The second step is the camera placement and adjustment According to the arranged measurement scene, determine the placement distance and angle between the two cameras, and adjust the focal length to make the cameras focus at infinity; The third step is for the left camera to collect cross calibration target images Control the left camera to collect cross calibration target images at different positions and poses; The fourth step is the decoding of cross calibration target images Extract the coordinates of the fiducial points in the cross calibration target images collected in the third step, and then perform decoding to obtain the fiducial point IDs; The fifth step is the orientation of cross calibration target images Select two cross calibration target images decoded in the fourth step for relative orientation; After successful orientation, perform absolute orientation on the other images decoded in the fourth step, and then obtain the exterior orientation elements of all images and the 3D coordinates of the fiducial points; The sixth step is the optimization of camera internal parameters Taking the reprojection error as the objective equation, use the camera internal parameters, the exterior orientation elements of the cross calibration target images obtained in the fifth step, and the 3D coordinates of the fiducial points as the parameters to be optimized for overall bundle adjustment to obtain the optimal camera internal parameters; The seventh step is the calibration of the right camera internal parameters According to the method from the third step to the sixth step, calibrate the internal parameters of the right camera; The eighth step is for the left and right cameras to synchronously collect one frame of measurement scene images Control the left and right cameras to synchronously collect one frame of the measurement scene images arranged in the first step; The ninth step is the fiducial point detection Detect the central coordinates of the circular non-coded fiducial points in the measurement scene images collected by the left and right cameras; The tenth step is the fiducial point matching Adopt the shape context algorithm to achieve the fast matching of the circular non-coded fiducial points detected in the ninth step; The eleventh step is the solution of the initial values of the external parameters Take the left camera coordinate system as the reference coordinate system, that is, the rotation matrix in the external parameters of the left camera is the identity matrix and the elements of the translation matrix are all 0. According to the fiducial point matching results in the tenth step, use the relative orientation algorithm to obtain the initial values of the external parameters of the right camera; then, reconstruct the 3D coordinates of the circular non-coded fiducial points within the measurement scene; The twelfth step is the total station measurement of the point distance Use the total station to measure the distance between two relatively far circular non-coded fiducial points within the measurement scene reconstructed in the eleventh step, where the distance between the two measured circular non-coded fiducial points should be more than half of the dimension in the length direction of the measurement scene; The thirteenth step is the correction and optimization of the camera external parameters Take the point distance measured in the twelfth step as the scale, calculate the scaling factor, correct the external parameters of the right camera solved in the eleventh step, and use the bundle adjustment method to optimize the external parameters of the right camera; The fourteenth step is to export the calibration data Export the internal and external parameters of the left and right cameras to complete the camera calibration.

2. The method for on-site rapid calibration of a large field-of-view camera applicable to complex working conditions according to claim 1, characterized in that, In the second step, the method for placing and adjusting the cameras is as follows: the included angle θ between the two cameras is in the range of 20° to 30°, that is, 20° < θ < 30°, and the measurement scene size is V x ×V y , with the unit of mm; the camera resolution is W × H; the pixel size is px, with the unit of μm / pixel; the focal length of the camera lens is f, with the unit of mm; the calculation formula for the distance D from the two cameras to the object to be measured is: D ≈ Vx × f / (W × px), with the unit of mm; the calculation formula for the placement distance L between the two cameras is: L ≈ 2 × D × sin(θ / 2); after the cameras are placed, adjust the focal length, that is, rotate the focus ring to the infinity mark [∞] on the lens, so that both cameras can clearly image within a distance from the hyperfocal point to the object to be measured.

3. The method for on-site rapid calibration of a large field-of-view camera applicable to complex working conditions according to claim 1, characterized in that, In the third step, the cross calibration target is a cross structure with four equal-length arms composed of profile I and profile II, which are equal in length and perpendicular to each other. On the center line of one of the arms belonging to profile I, N marking points are printed or pasted at equal intervals, where N≥4. On the center lines of the other three arms, N + 1 marking points are printed or pasted at equal intervals. The cross ratio of every 4 adjacent marking points on each arm of the target is equal to 4 / 3. The marking points printed or pasted on each profile of the target are all circular non-coded marking points of the same size. The marking points printed or pasted on each profile of the target are required to be collinear.

4. The method for on-site rapid calibration of a large field-of-view camera applicable to complex working conditions according to claim 3, characterized in that, The coding rule of the cross calibration target is as follows: For profile I with 2N + 1 marking points printed or pasted, in the direction from the starting end to the terminal end, the IDs of the marking points are sequentially assigned from 0 to 2N. For profile II with 2N + 2 marking points printed or pasted, in the direction from the starting end to the terminal end, the IDs of the marking points are sequentially assigned from 2N + 1 to 4N + 2. The starting end and the terminal end of profile I and profile II are defined as follows: The starting end of profile I is defined as the end where N marking points are printed or pasted on the arm, and the terminal end is defined as the end where N + 1 marking points are printed or pasted on the arm. When profile I is rotated counterclockwise by 90° to coincide with profile II, the end of profile II that coincides with the starting end of profile I is defined as the starting end of profile II, and the end that coincides with the terminal end of profile I is defined as the terminal end of profile II.

5. The on-site rapid calibration method for a large field-of-view camera applicable to complex working conditions according to claim 3 or 4, characterized in that, In the fourth step, the cross calibration target image decoding process is as follows: 1) Detection of circular non-coded marking points Extract the center coordinates of the marking points in the cross calibration target image collected in the third step. 2) Line fitting Use the RANSAC algorithm to sequentially fit two lines to the extracted center coordinates and mark the inliers corresponding to each line. 3) Line segmentation Calculate the intersection point of the two fitted lines and use it as the segmentation point to divide each line into two segments. Divide the inliers of each marked line into two groups, corresponding to the two segments after segmentation respectively. 4) Cross ratio invariance test Calculate the cross ratio of the inliers of each segment of the divided line and mark the inliers that meet the cross ratio invariance condition as the marking points on the divided line. 5) Verification of the number of marking points Check whether the number of marking points on the four segments of the divided line after the cross ratio invariance test is the same as the number of marking points printed or pasted on the four arms of the cross calibration target in the third step: If the number of marking points on one of the divided lines is N and the number of marking points on the other three divided lines is N + 1, it is determined to be qualified. 6) ID assignment of marking points If the verification of the number of marking points is qualified, then sequentially assign IDs to the marking points on the four segments of the divided line.

6. The on-site rapid calibration method for a large field-of-view camera applicable to complex working conditions according to claim 5, characterized in that, The detection of circular non-coded marking points is implemented by the following real-time detection algorithm for circular non-coded marking points: First, initialize on the CPU side and the GPU side, that is, allocate enough pinned memory on the CPU side to store the target image collected in the third step, and allocate enough memory areas in the global memory on the GPU side for data storage during the GPU-side processing. Secondly, upload the target image collected in the third step from the CPU side to the GPU side. Asynchronous transmission is adopted during the transmission. After the image transmission is completed, start the binarization process on the GPU side. After the binarization is completed, use the two-step method to perform connected component labeling and analysis on the GPU side, and on this basis, extract the integer pixel edges of the circular non-coded fiducial points; Next, on the GPU side, extract the sub-pixel edges based on the spatial moment method, and use the least squares ellipse fitting algorithm to obtain the center coordinates of the circular non-coded fiducial points; Finally, asynchronously transmit the detection result from the GPU side to the CPU side to complete the real-time detection of the fiducial points in a calibrated target image collected in the third step.

7. The on-site rapid calibration method for a large field-of-view camera applicable to complex working conditions according to claim 5, characterized in that, The method for cross-ratio invariance test is as follows: For the inliers of a segmented line, if the number is less than N, it is determined that all inliers do not meet the cross-ratio invariance condition; if the number is equal to or more than N, calculate the distance from each inlier to the intersection point to be found, and sort them in ascending order of distance; Take the first 4 inliers after sorting, calculate the cross-ratio λ. If it meets the cross-ratio invariance condition: 4 / 3 - ε ≤ λ ≤ 4 / 3 + ε, it is determined that the first 4 inliers all meet the invariance condition, and mark the first 4 inliers as the fiducial points on the segmented line; For the 5th inlier after sorting, take the first three inliers adjacent to it, that is, the 4th, 3rd, and 2nd inliers after sorting, to form a set of adjacent 4 inliers, calculate the cross-ratio λ. If it meets the cross-ratio invariance condition: 4 / 3 - ε ≤ λ ≤ 4 / 3 + ε, it is determined that the 5th inlier after sorting meets the invariance condition, and mark it as the fiducial point on the segmented line; Finally, sequentially check the other inliers after sorting according to the determination method of the 5th inlier; Among them, ε represents the error factor, and the value range of ε is: ε = 0.02 - 0.

08.

8. The on-site rapid calibration method for a large field-of-view camera applicable to complex working conditions according to claim 5, characterized in that, The method for fiducial point ID assignment is as follows: For a segmented line with N fiducial points, calculate the distance from each fiducial point on it to the intersection point to be found, and sort the fiducial points in descending order of distance. After sorting, assign the IDs of the fiducial points as 0 to N - 1 in sequence; For another segmented line collinear with a segmented line containing N fiducial points, calculate the distance from each fiducial point on it to the intersection point to be found, and sort the fiducial points in ascending order of distance. After sorting, assign the IDs of the fiducial points as N to 2N in sequence; Secondly, construct a vector with the starting point being the fiducial point with ID = 0 and the ending point being the fiducial point with ID = 2N, and use this vector as the reference vector. For the other two segmented straight lines, calculate the distances between the fiducial points on each straight line and the intersection point to be found respectively, and select the fiducial point with the farthest distance on each straight line. Connect it with the fiducial point with ID = 0 to form a vector, where the fiducial point with ID = 0 is the starting point of the vector, and then take the cross product with the constructed reference vector. If the cross product is positive, sort all the fiducial points on this segmented straight line in descending order according to the calculated distances, and assign IDs to the fiducial points in sequence from 2N + 1 to 3N + 1 after sorting; if the cross product is negative, sort all the fiducial points on this segmented straight line in ascending order according to the calculated distances, and assign IDs to the fiducial points in sequence from 3N + 2 to 4N + 2 after sorting.

9. The on-site rapid calibration method for a large field-of-view camera applicable to complex working conditions according to claim 5, characterized in that, In the fifth step, during image orientation, calculate the angle ρ between the two images along the optical axis direction according to the rotation matrix. If the angle ρ > 20°, the relative orientation is successful. Otherwise, the orientation fails. Replace with another two images for relative orientation until it is successful. After the relative orientation is successful, calculate the three-dimensional coordinates of the fiducial points in the two images, use these fiducial points as control points, and use the pyramid method to perform absolute orientation on other images to obtain the exterior orientation elements of other images.

10. The on-site rapid calibration method for a large field-of-view camera applicable to complex working conditions according to claim 5, characterized in that, In the thirteenth step, the method for camera external parameter calibration and optimization is as follows: Take the point distance measured by the total station in the twelfth step as the scale, calculate the scaling factor, scale the three-dimensional coordinates of all fiducial points in the measurement scene and the initial value of the external parameters of the right camera solved in the eleventh step to obtain the calibrated external parameters of the right camera; finally, use the bundle adjustment method, take the reprojection error as the optimization target, and iteratively solve the optimal camera external parameters.

Images