A feature-based joint calibration method for calibrating multiple cameras for cooperative shooting

Abstract

The application relates to the technical field of automobile auxiliary driving, in particular to a kind of feature joint calibration multiple camera cooperative shooting external parameter calibration method based on, applied to the joint calibration of vehicle-mounted ring view system and unmanned aerial vehicle overhead camera, first, the image collected by vehicle camera is aligned with the overhead view photographed by unmanned aerial vehicle from above simultaneously to the same natural feature, second, ORB is used for feature point extraction and rough matching, then fine matching is realized through LK, finally, the accurate homography transformation from vehicle-mounted perspective to real vertical overhead perspective is calculated by using RANSAC to remove abnormal matching, the feature joint calibration multiple camera cooperative shooting external parameter calibration method based on the application not only gets rid of the dependence on specific calibration scene and artificial auxiliary tool, but also improves the flexibility and precision of calibration calculation by fully utilizing natural features, and the geometric accuracy of bird's-eye view image generated by 360-degree ring view image system is significantly optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automobile auxiliary driving, and particularly relates to a multi-camera cooperative shooting extrinsic parameter calibration method based on feature joint calibration. BACKGROUND

[0002] The automobile 360° surround view image system is an advanced driving assistance technology. It captures the environment around the vehicle through multiple wide-angle cameras installed around the vehicle, processes and fuses these images, and generates a complete overhead panoramic image, providing the driver with a visual field without dead angles. A typical surround view system consists of four fisheye cameras, which are generally installed on the front, rear, left and right of the vehicle. The central control unit splices and fuses the images to generate a 360-degree overhead panoramic image in real time. This technology has been widely used in parking assistance and low-speed obstacle avoidance scenarios, greatly improving driving safety.

[0003] In the process of implementing the 360° surround view image system, the overall process generally includes image acquisition and preprocessing, intrinsic parameter calibration, extrinsic parameter calibration and image splicing. First, the intrinsic parameter calibration of each camera is required, that is, the focal length, principal point position and lens distortion parameters are corrected. This process is relatively mature, and the Zhang Zhengyou chessboard calibration method is usually used to obtain the focal length, principal point and distortion coefficient of the camera. The intrinsic parameters of the fisheye camera of the commercial surround view system are generally fixed at the factory.

[0004] In contrast, extrinsic parameter calibration, i.e., determining the relative positions and attitudes of each camera, has long been a key and challenging aspect of accurate imaging in surround-view systems. Traditional methods typically require placing calibration patterns around the vehicle and calculating the camera's attitude relative to the vehicle's coordinate system by photographing the known reference points and their image positions in each camera. A common practice is to place four large checkerboard patterns or calibration boards on the ground around the car (located near the front left, front right, rear left, and rear right of the vehicle, respectively), their relative positions pre-measured and known. Each fisheye camera must be able to see at least two corner points of the calibration boards, and the projection matrix / homography transformation of that camera is calculated through the correspondence between image coordinates and coordinates on the bird's-eye view plane. This method requires acquiring a large number of images and undergoing multiple steps of data processing and optimization calculations. It also relies on manual assistance tools (such as checkerboard patterns or specific calibration boards) to acquire data under static conditions and requires the calibration boards to be arranged reasonably so that the fields of view of adjacent cameras have overlapping references for joint calibration. In practical applications, due to factors such as vibration and temperature changes during vehicle operation, the extrinsic parameters between cameras may drift. At this point, the original joint calibration results are no longer accurate, thus affecting the image stitching effect and the overall system performance. Once extrinsic parameter drift occurs, it is usually necessary to send the vehicle to a professional site for recalibration by engineers with specialized equipment and extensive experience. This is not only time-consuming and labor-intensive, but also increases vehicle maintenance and management costs, and makes it difficult to restore the system's accuracy in a timely manner in emergency situations.

[0005] Therefore, in the field of extrinsic parameter calibration for automotive 360° surround view imaging systems, existing technologies have the following main drawbacks:

[0006] 1. Traditional methods usually require a checkerboard, calibration board or specific pattern as calibration target, require vehicles to be parked in a special area, and are complicated and costly to operate;

[0007] 2. Using the calibration board requires precise placement and multi-angle shooting, and relies on manual adjustment of the camera position or the placement of the calibration board, which is time-consuming.

[0008] 3. Traditional extrinsic parameter calibration methods usually rely on an approximate projection model from a single vehicle-mounted viewpoint, which makes it difficult to accurately establish the geometric correspondence between the camera and the real ground plane, resulting in significant geometric distortion in the generated bird's-eye view. In view of the above, there is an urgent need to develop an extrinsic parameter calibration method based on feature joint calibration and collaborative shooting by multiple cameras to overcome the shortcomings in current practical applications. Summary of the Invention

[0009] The purpose of this invention is to provide a method for extrinsic parameter calibration based on feature joint calibration of multiple cameras for collaborative shooting, in order to solve the problems of extrinsic parameter calibration in 360° surround view imaging systems relying on manual auxiliary tools and the difficulty in accurately calibrating the geometric correspondence between vehicle-mounted cameras and the real ground plane.

[0010] To achieve the above object, the present application provides the following technical solutions:

[0011] A feature joint-based calibration method for multiple camera cooperative shooting extrinsic parameter calibration, applied to joint calibration of a vehicle-mounted surround view system and a UAV overhead camera, comprising the following steps:

[0012] S1, cooperative data acquisition: simultaneously shooting a ground scene containing the same natural features by at least one vehicle-mounted camera of the vehicle-mounted surround view system and an overhead camera carried by a UAV, to obtain vehicle-mounted camera images and overhead camera images;

[0013] S2, feature point extraction and coarse matching: extracting natural feature points from the vehicle-mounted camera images and the overhead camera images respectively and generating corresponding feature descriptors, preliminarily matching based on the similarity of the feature descriptors, and obtaining an initial feature point matching pair set;

[0014] S3, matching fine adjustment: optimizing the positions of the matching points in the initial feature point matching pair set, and obtaining fine-adjusted feature point matching pairs;

[0015] S4, extrinsic parameter calculation: based on the fine-adjusted feature point matching pairs, using a random sample consensus algorithm to eliminate abnormal matching points and calculating a homographic transformation matrix between the vehicle-mounted camera and the overhead camera, the homographic transformation matrix representing the projection transformation relationship from the vehicle-mounted camera perspective to the overhead camera perspective.

[0016] As a further scheme of the present application, after step S1 and before step S2, there is further:

[0017] S1a, image preprocessing: using pre-calibrated vehicle-mounted camera intrinsic parameters and overhead camera intrinsic parameters to respectively correct the distortion of the vehicle-mounted camera images and the overhead camera images; the intrinsic parameters include focal length, principal point coordinates, radial distortion coefficients and tangential distortion coefficients.

[0018] As a further scheme of the present application, step S2 specifically comprises:

[0019] S2a, feature point extraction: using an ORB feature detection algorithm to extract natural feature points from the vehicle-mounted camera images and the overhead camera images respectively;

[0020] S2b, descriptor generation: generating an ORB binary descriptor for each natural feature point extracted;

[0021] S2c, coarse matching: calculating the Hamming distance between the feature point descriptors of the vehicle-mounted camera images and the feature point descriptors of the overhead camera images, sorting them from small to large according to the Hamming distance, and selecting the top N best matching pairs as the initial feature point matching pair set.

[0022] As a further scheme of the present application: step S3 is specifically:

[0023] The LK optical flow method is used to optimize the position of each matching point in the initial feature point matching pair set;

[0024] Wherein, the vehicle-mounted camera image is the source image, the feature points thereon are the initial positions, the overhead camera image is the target image, and the preliminary matching points thereon are the positions to be optimized; the matching point positions on the target image are optimized by minimizing the image brightness difference in the local window to obtain the fine-tuned feature point matching pairs.

[0025] As a further scheme of the present application: step S4 specifically includes:

[0026] S4a, model fitting and inlier selection: a minimum point set is randomly selected from the fine-tuned feature point matching pairs, and a candidate homography matrix is calculated; all vehicle-mounted camera image feature points are projected to the overhead camera image coordinate system using the candidate homography matrix, and the re-projection error between the projected points and the actual matching points is calculated; the matching points with a re-projection error less than a preset threshold are marked as inliers;

[0027] S4b, optimal model selection: step S4a is repeated multiple times, and the candidate homography matrix with the most inliers is selected as the optimal homography transformation model;

[0028] S4c, model optimization: using all inliers corresponding to the optimal homography transformation model, the direct linear transformation method is used for re-fitting to obtain an optimized homography transformation matrix.

[0029] As a further scheme of the present application: the minimum point set contains 4 matching pairs; the candidate homography matrix is calculated by solving a linear equation set using the direct linear transformation method.

[0030] As a further scheme of the present application: the natural features include building outlines, road signs, or vegetation edges.

[0031] As a further scheme of the present application: further comprising: using the homography transformation matrix to project the vehicle-mounted camera image to the overhead camera view plane to generate a bird's eye view through image stitching;

[0032] The calibration effect is verified by comparing the stitched bird's eye view with the real overhead camera image, or by observing the geometric shape, position, and scale consistency of a preset reference object in the stitched image.

[0033] Compared with the prior art, the present application has the following advantages:

[0034] The application has practical significance in the external parameter calibration of the automobile 360° surround view image system. First, the joint calibration of the vehicle-mounted view angle and the UAV overhead view angle is performed by using natural features, which eliminates the dependence on traditional calibration tools (such as a checkerboard calibration board), makes the calibration process more flexible, reduces the field preparation work, and improves the adaptability of the system. Secondly, the application adopts a multi-step matching optimization process, including ORB coarse matching, LK optical flow fine adjustment and RANSAC algorithm, which significantly improves the matching accuracy and effectively eliminates abnormal matching points, ensuring the high precision and stability of the external parameter calibration result. Experimental results show that the method can improve the geometric accuracy of the bird's-eye view generated by the surround view system. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The figure is a data acquisition and preprocessing schematic diagram in the embodiment of the application.

[0036] Figure 2 The figure is an ORB feature detection module and ORB coarse matching schematic diagram in the embodiment of the application.

[0037] Figure 3 The figure is a homography matrix calculation schematic diagram in the embodiment of the application.

[0038] Figure 4 The figure is an image stitching checking schematic diagram in the embodiment of the application. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0040] The specific implementation of the application will be described in detail below in combination with specific embodiments.

[0041] Please refer to Figures 1-4The embodiment of the application provides a multi-camera cooperative shooting extrinsic parameter calibration method based on feature joint calibration, first, a ground scene is captured by a UAV camera and a vehicle-mounted surround-view camera simultaneously, natural feature points (such as building outlines, road signs or vegetation edges) in the ground scene are extracted by using a Features from Accelerated Segment Test (FAST) algorithm and a Binary Robust Independent Elementary Features (BRIEF) algorithm detection technology (Oriented FAST and Rotated BRIEF, ORB), and ORB coarse matching is performed, and the first N best matching pairs are selected according to a matching distance order. Subsequently, a Lucas-Kanade Optical Flow (LK) method is applied to fine-tune the coarse matching result, and a more accurate feature point correspondence relationship is obtained. Finally, a RANdom SAmple Consensus (RANSAC) algorithm is used to remove abnormal matching and calculate a homography conversion matrix of the vehicle-mounted camera to the UAV camera, so that accurate conversion from a vehicle-mounted perspective to a bird's eye view is realized.

[0042] Embodiment 1: core method flow;

[0043] Step 1, data acquisition and pretreatment: including a data acquisition module and an image pretreatment module;

[0044] The data acquisition module is responsible for simultaneously shooting a ground scene by a UAV camera and a vehicle-mounted surround-view camera, and acquiring original image data of the two; the UAV camera provides a bird's eye view, and the vehicle-mounted surround-view camera provides a surround view around the vehicle, and the two provide original data sets for other modules;

[0045] The image pretreatment module is responsible for carrying out distortion removal on the acquired original image, correcting the image by using known camera intrinsic parameters (such as focal length, principal point coordinates) and distortion coefficients (such as radial distortion and tangential distortion), and ensuring geometric accuracy of subsequent feature detection and matching;

[0046] Step 2, feature detection and coarse matching: including an ORB feature detection module and an ORB coarse matching module;

[0047] The ORB feature detection module is responsible for extracting natural feature points in the UAV bird's eye view image and the vehicle-mounted surround-view image and generating corresponding descriptors;

[0048] The ORB coarse matching module is responsible for finding matching feature point pairs in two images (a vehicle-mounted camera image and a UAV bird's eye view image);

[0049] Step 3, fine matching: specifically includes an LK optical flow module, which is responsible for optimizing the feature point matching pairs obtained by coarse matching to make the position more accurate;

[0050] Step 4, homography matrix calculation: specifically includes an RANSAC algorithm module, which uses the RANSAC algorithm to remove abnormal matching points (outliers) from the fine-tuned feature point matching pairs, and calculates the homography transformation matrix from the vehicle-mounted camera to the UAV camera;

[0051] Step 5, image stitching and qualitative evaluation: after successfully calculating the homography matrix from the vehicle-mounted camera to the UAV camera, the matrix is used for image stitching to fuse multiple vehicle-mounted camera images into a seamless bird's eye view, providing an intuitive overhead perspective for the car 360° surround view system.

[0052] Embodiment 2: image preprocessing;

[0053] In step 1, the image processing process of each camera is as follows:

[0054] ;

[0055] wherein, is the two-dimensional image pixel coordinate under the condition of no distortion, is the three-dimensional world coordinate; the camera intrinsic parameter defines the projection relationship from the three-dimensional point to the two-dimensional image plane, and the intrinsic matrix is used to represent:

[0056] ;

[0057] wherein, , is the focal length (pixel unit), , is the principal point coordinate (image center); this formula describes the ideal projection process, but the actual image needs to be further corrected due to lens distortion. Distortion correction maps the pixel coordinates of the distorted image back to the non-distorted coordinates through the radial distortion and tangential distortion models. Taking normalized coordinates as an example, the distortion correction process is:

[0058] ;

[0059] ;

[0060] wherein, is the normalized coordinate without distortion, ; is the radial distortion coefficient, is the tangential distortion coefficient; is the normalized coordinate after distortion; is solved by inverse transformation , and then converted to pixel coordinates by intrinsic conversion:

[0061] ;

[0062] The image pre-processing module ensures the geometric accuracy of the images by intrinsic projection and distortion correction.

[0063] Example 3: ORB feature detection and coarse matching

[0064] In step 2, ORB feature detection is performed by using the improved FAST algorithm to detect the corner points in the image. The FAST algorithm judges whether a pixel point is a corner point by comparing its intensity with the intensities of its surrounding pixels. For a pixel point , its 16 surrounding pixels (usually distributed on a circle with a radius of 3) are checked. If there are consecutive pixels (usually ) whose intensities are greater or less than the intensity of by a threshold , then is considered as a corner point.

[0065] To make ORB rotation-invariant, ORB adjusts the orientation of the descriptor by calculating the orientation of the corner point. The orientation of the corner point is calculated by the centroid of the image patch:

[0066] ;

[0067] where and are the first-order moments of the image patch:

[0068] ;

[0069] Here denotes the intensity of pixel in the image patch;

[0070] Then, a binary descriptor is generated using the BRIEF algorithm. BRIEF generates a binary string as the descriptor by comparing the intensities of a predefined pair of pixels in the image. The descriptor generation process is described as follows:

[0071] First, a pair of pixels around the feature point is selected, and their intensities are compared:

[0072] ;

[0073] The descriptor is a binary string of bits:​

[0074] ;

[0075] To make the descriptor rotation-invariant, ORB rotates the keypoint according to its orientation Rotate the position of the pixel pair:

[0076] ;

[0077] where, is the rotation matrix:

[0078] ;

[0079] ORB matches the feature points in two images, using the Fast Library for Approximate Nearest Neighbors (FLANN) to compute the Hamming distance between each pair of descriptors. The smaller the Hamming distance, the more similar the two descriptors. The Hamming distance between two binary descriptors and is defined as:

[0080] ;

[0081] where, is the XOR operation, is the number of bits in the descriptor; the result of the XOR operation is 1 if the two bits are different, and 0 if the two bits are the same, so the Hamming distance represents the number of different bits between the two descriptors.

[0082] Subsequently, all the descriptor pairs are sorted by Hamming distance from small to large, and the top best matching pairs are selected. The value of can be adjusted according to the image size and the number of feature points, determined by empirical value or cross-validation (when the image is 1920x1080 and the number of feature points is about 3000, the empirical value is , where is the total number of feature points in the source image; in 4K images or scenarios with higher accuracy requirements, the scaling factor can be adjusted from 0.2 to 0.3-0.4). Finally, a set of initial feature point matching pairs is obtained, which serves as the input for subsequent fine tuning (such as homography matrix calculation).

[0083] Example 4: LK optical flow fine-tuning matching

[0084] In step 3, for the source image (shot by a vehicle-mounted camera) and the target image ​, the corresponding preliminary matching point in the target image is The LK optical flow method assumes that the intensity of a feature point in an image remains constant over a short time and that the motion vectors of neighboring pixels are similar. Based on these assumptions, the LK optical flow method adjusts the preliminary point in the target image to a more accurate position .

[0085] First, a local window is defined around the feature point in the source image (the window radius is empirically set to be 15-31 pixels, which can be adaptively adjusted according to the resolution). For the target image , the gradient of each pixel in the local window in the horizontal direction and the vertical direction is calculated:

[0086] ;

[0087] Subsequently, the intensity difference of the corresponding positions in the local window in the source image and the target image is calculated:

[0088] ;

[0089] wherein reflects the inconsistency of the gray scale of the two images in the local area, which is usually caused by the displacement of the feature point; the LK optical flow method estimates the displacement by minimizing this difference.

[0090] Using the gradient and the intensity difference in the local window , a linear equation system can be constructed to solve the displacement vector . The optical flow equation is constructed by the matrix and the vector :

[0091] ;

[0092] wherein is a 2x2 symmetric matrix representing the distribution characteristics of the gradient; the vector represents the weighted sum of the gradient and the intensity difference:

[0093] ;

[0094] ; ​

[0095] Calculate displacement vector by solving linear equations , and apply the calculated to the initial points of the target image :

[0096] ;

[0097] Finally, through iterative optimization, a higher-precision correspondence is obtained, thereby improving the matching quality.

[0098] Example 5: RANSAC homography matrix calculation

[0099] In step 4, the calculation formula of the homography matrix H is:

[0100] ;

[0101] wherein, is the point coordinate in the vehicle-mounted camera image, is the corresponding point coordinate in the UAV camera image, is a 3x3 homography matrix.

[0102] RANSAC randomly samples a minimum point set (4 pairs of points) multiple times, fits a homography model, and counts the number of inliers, finally selects the model with the most inliers as the result. From the fine-tuned feature point matching pairs, 4 pairs of points are randomly selected, and a candidate homography matrix is calculated using the selected 4 pairs of matching points through the direct linear transformation (DLT) method. The specific calculation process is as follows:

[0103] For each pair of matching points, let the feature point coordinate in the vehicle-mounted camera image be , and the corresponding point coordinate in the UAV camera image be According to the mathematical definition of homographic transformation, the following relationship is satisfied:

[0104] ;

[0105] ;

[0106] To eliminate the nonlinear influence of the denominator, cross-multiplication is used to convert it to a linear equation, which is:

[0107] ;

[0108] For 4 pairs of matching points, an 8x8 linear equation set can be constructed. The singular value decomposition (SVD) method is used to solve the equation set, and the vector form of the candidate homography matrix is obtained and normalized.

[0109] The calculated candidate matrix , all feature points in the vehicle-mounted camera image are transformed into the UAV camera image coordinate system to generate predicted point coordinates . The re-projection error between the predicted point and the actual matching point is calculated and defined as:

[0110] ;

[0111] If the re-projection error of a certain feature point is less than a preset threshold (2 pixels are set according to the image resolution and camera calibration accuracy experience, about 0.06% of the diagonal line of a 1920x1080 image), the point is marked as an inlier, otherwise it is considered as an outlier. Through repeated iteration, the candidate with the largest number of inliers is selected as the preliminary best model. Further, using all inliers corresponding to the best model, an optimized homography matrix is fitted through the DLT method, thereby improving the transformation accuracy and stability.

[0112] Example 6: Calibration effect verification

[0113] In step 5, the homography matrix is used to project the image of each vehicle-mounted camera to the top-down plane of the UAV camera. The specific calculation process is as follows:

[0114] For a pixel point in the vehicle-mounted camera image, the corresponding homogeneous coordinates are obtained through matrix transformation, wherein

[0115] ;

[0116] Then the actual pixel coordinates are obtained by normalization:

[0117] ;

[0118] Then check whether the overall stitching image is smooth and has no obvious seams. Compare the stitching image with the real UAV top-down image to verify whether the shapes, positions and proportions of objects (such as vehicles, pedestrians and stop lines, etc.) are accurate and consistent.

[0119] In order to check the accuracy of image stitching, the vehicle can be parked on a large chessboard, and by observing the shape, position and proportion of the chessboard in the stitched image, the stitching effect can be evaluated whether it is ideal.

[0120] To sum up, first, the present application avoids the dependence on traditional artificial calibration tools (such as checkerboard calibration board) by using natural features (such as building contours, road signs, etc.) for joint calibration of vehicle-mounted cameras and overhead cameras of unmanned aerial vehicles. Cross-platform accurate perspective conversion is achieved, improving the flexibility and applicability of the external parameter calibration, which can effectively reduce the manual intervention and scene restrictions. Second, the present application uses the ORB feature detection algorithm for coarse matching, fine-tunes the matching results by the LK optical flow method, and uses the RANSAC algorithm to eliminate abnormal matching points to calculate the homography matrix. Through this multi-stage matching optimization process, not only the matching accuracy is improved, but also the robustness to noise and outliers in the calibration process is enhanced, ensuring high-precision external parameter calculation. Therefore, the present application realizes accurate bird's-eye view conversion from the vehicle-mounted perspective to the real vertical overhead perspective, which helps to improve the geometric accuracy of the 360° surround view image system.

[0121] It should be noted that in the present application, it should be understood that although the present application is described in terms of embodiments, each embodiment does not contain only one independent technical solution, and the description manner of the specification is only for the sake of clarity, and those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be properly combined to form other embodiments that those skilled in the art can understand.

Claims

1. A method for calibrating extrinsic parameters of multiple cameras based on feature joint calibration, characterized in that, The application is applied to joint calibration of a vehicle-mounted surround view system and a UAV overhead camera, and includes the following steps: S1, cooperative data acquisition: simultaneously shooting a ground scene containing the same natural features by at least one vehicle-mounted camera of a vehicle-mounted surround view system and an overhead camera carried by a UAV, and acquiring vehicle-mounted camera images and overhead camera images; S2, feature point extraction and coarse matching: extracting natural feature points from the vehicle-mounted camera images and the overhead camera images respectively and generating corresponding feature descriptors, preliminarily matching based on the similarity of the feature descriptors, and obtaining an initial feature point matching pair set; S3, matching fine adjustment: optimizing the matching point positions in the initial feature point matching pair set, and obtaining fine-adjusted feature point matching pairs; S4, external parameter calculation: based on the fine-adjusted feature point matching pairs, using a random sample consensus algorithm to remove abnormal matching points and calculating a homographic transformation matrix between the vehicle-mounted camera and the overhead camera, wherein the homographic transformation matrix represents the projection transformation relationship from the vehicle-mounted camera perspective to the overhead camera perspective; After step S1 and before step S2, the following steps are further included: S1a, image preprocessing: using pre-calibrated vehicle-mounted camera internal parameters and overhead camera internal parameters to respectively correct the distortion of the vehicle-mounted camera images and the overhead camera images; the internal parameters include focal length, principal point coordinates, radial distortion coefficients and tangential distortion coefficients; Step S2 specifically includes: S2a, feature point extraction: using an ORB feature detection algorithm to extract natural feature points from the vehicle-mounted camera images and the overhead camera images respectively; S2b, descriptor generation: generating an ORB binary descriptor for each natural feature point extracted; S2c, coarse matching: calculating the Hamming distance between the feature point descriptors of the vehicle-mounted camera images and the feature point descriptors of the overhead camera images, sorting the Hamming distances from small to large, and selecting the top N best matching pairs as the initial feature point matching pair set; Step S3 specifically includes: using an LK optical flow method to optimize the positions of each matching point in the initial feature point matching pair set; wherein the vehicle-mounted camera image is taken as a source image, the feature points thereon are taken as initial positions, the overhead camera image is taken as a target image, and the preliminary matching points thereon are taken as positions to be optimized; the matching point positions on the target image are optimized by minimizing the image brightness difference in a local window, and fine-adjusted feature point matching pairs are obtained; Step S4 specifically includes: S4a, model fitting and inlier selection: randomly selecting a minimum point set from the fine-adjusted feature point matching pairs, calculating a candidate homographic matrix; projecting all vehicle-mounted camera image feature points to the overhead camera image coordinate system using the candidate homographic matrix, calculating the re-projection error between the projected points and the actual matching points, and marking the matching points with a re-projection error less than a preset threshold as inliers; S4b, optimal model selection: repeating step S4a multiple times, and selecting the candidate homographic matrix with the most inliers as the optimal homographic transformation model; S4c, model optimization: using all inliers corresponding to the optimal homographic transformation model to re-fit an optimized homographic transformation matrix through a direct linear transformation method.

2. The method of claim 1, wherein, The minimum point set contains 4 pairs of matching points; and the calculation of the candidate homography matrix is achieved by solving linear equations using a direct linear transformation method. 3.The method of claim 1, wherein, The natural features include building outlines, road signs, or vegetation edges.

4. The method of claim 1, wherein, The method further comprises: projecting the vehicle camera image to a top-view camera visual angle plane using the homographic transformation matrix, and generating a bird's-eye view by image stitching. The calibration effect is verified by comparing the stitched bird's-eye view with a real top-view camera image, or by observing the geometric shape, position, and scale consistency of a preset reference object in the stitched image.

Images