A camera automatic calibration method for a highway scene

Abstract

This invention belongs to the field of computer vision and intelligent transportation technology. It discloses an automated camera calibration method for highway scenarios. Spatial coordinate data is acquired by combining static point marking with dynamic trajectory tracking, and the timestamps embedded in the video stream are used to achieve precise matching between world coordinates and pixel coordinates. This method introduces a height vector compensation mechanism to eliminate antenna elevation deviation and integrates YOLO-pose keypoint detection and ByteTrack target tracking technologies to achieve automated extraction of grounding point pixel coordinates and cross-frame trajectory association. Simultaneously, a region-adaptive calibration strategy is implemented based on road geometric features, and a feature-matching-based calibration parameter drift determination mechanism is established. This invention reduces manual workload and safety risks, improves calibration efficiency and accuracy, and provides a stable and reliable coordinate mapping foundation for traffic monitoring, event analysis, and intelligent management in highway scenarios.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and intelligent transportation technology, and more specifically, to an automated camera calibration method for highway scenarios. Background Technology

[0002] Currently, camera calibration technology in complex traffic scenarios such as highways faces numerous unresolved technical challenges. The unique characteristics of highway scenarios—long viewing distances and wide fields of view—lead to significant resolution differences between near and far regions in the image. Traditional single-mapping models struggle to simultaneously guarantee accurate positioning of both near lane lines and distant vehicles. Surveillance cameras are typically mounted on high poles or overpasses, employing steep downward angles for shooting, resulting in severe perspective and radial distortion in the images. This makes calibration methods suitable for controlled environments such as standard checkerboard patterns unsuitable. In practical applications, the core driving areas of highways are high-risk, restricted areas. Traditional calibration methods requiring manual entry pose serious safety hazards and often necessitate temporary lane closures, disrupting traffic flow. Cameras exposed to complex weather environments are frequently affected by wind vibrations, diurnal temperature variations, and rain and snow erosion, causing camera parameters to drift. Existing technologies lack effective real-time monitoring and early warning mechanisms, resulting in the failure to detect calibration failures in a timely manner, thus impacting the accuracy of critical applications such as traffic incident detection and vehicle speed measurement. Meanwhile, the geometric shapes of different road sections vary significantly in actual engineering projects. For example, tunnel entrances and exits, interchanges, curves, and straight sections differ greatly, making it difficult for a unified calibration model to adapt to this diversity. Furthermore, the stability of image feature extraction under different weather and lighting conditions is insufficient, causing feature-matching-based calibration methods to perform poorly in complex environments such as rain, fog, and nighttime. In addition, the massive number of cameras in highway monitoring systems requires efficient calibration solutions, but existing technologies often require individual calibration by professionals, which is time-consuming, labor-intensive, and difficult to deploy on a large scale.

[0003] In view of this, the present invention proposes an automated camera calibration method for highway scenarios to solve the above problems. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: an automated camera calibration method for highway scenarios, comprising: Acquire spatial coordinate data and synchronous video data of the monitoring area in a highway scenario; Time matching is performed based on the acquisition timestamp of spatial coordinate data and the frame timestamp in synchronized video data to establish a set of calibration point pairs containing world coordinates and pixel coordinates; Based on the geometric distribution characteristics of roads within the monitoring area, the calibration point pair set is divided into regions to obtain multiple calibration sub-regions and their corresponding sub-region calibration point pairs; For each calibration sub-region, an adaptive homography matrix is ​​calculated for the calibration point pairs to obtain the pixel-world coordinate mapping relationship of each calibration sub-region; the generalization ability of the mapping relationship is evaluated based on cross-validation. Based on the feature matching relationship between the currently acquired image and the reference image, the drift amount of the camera calibration parameters is obtained; the calibration validity is determined based on the drift amount.

[0005] The technical effects and advantages of the automated camera calibration method for highway scenarios proposed in this invention are as follows: This invention, through a precise calibration strategy, enables the system to achieve high-precision perception of vehicle position, speed, and trajectory in complex and ever-changing highway conditions, enhancing traffic incident identification and analysis capabilities. The regional adaptive mechanism ensures consistent calibration results across the entire field of view, obtaining reliable coordinate mapping relationships whether it's near the road shoulder or in the far lane. This invention improves the safety of highway monitoring, avoiding the risks of frequent personnel entry into dangerous areas required by traditional methods, while minimizing interference with normal traffic flow. This invention uses YOLO-pose to precisely locate the grounding point, solving the projection deviation caused by traditional center point annotation; combined with ByteTrack target tracking technology, it simplifies the annotation of massive discrete frames to ID-based trajectory filtering, greatly improving calibration efficiency. Regular monitoring and judgment of calibration parameter drift enables effective control of system performance status, assisting subsequent calibration and maintenance decisions, significantly extending the calibration validity period, reducing the need for frequent maintenance and calibration, and saving operation and maintenance costs and human resources. Even under extreme weather and lighting conditions, this invention can still provide stable and reliable calibration results, providing a solid foundation for all-weather, all-time highway monitoring. Attached Figure Description

[0006] Figure 1 This is a schematic diagram of an automated camera calibration method for highway scenarios according to the present invention. Detailed Implementation

[0007] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0008] Please see Figure 1 In this embodiment of the invention, the detailed implementation steps of an automated camera calibration method for highway scenarios include: First, spatial coordinate data and synchronous video data of the monitoring area in the highway scenario are acquired. Spatial coordinate data includes precise geographical location information of road feature points within the monitoring area, collected using high-precision RTK measurement equipment. Synchronous video data is collected by fixedly installed surveillance cameras, recording video images within the same time period. These two sets of data constitute the basic input for camera calibration, providing the necessary basis for subsequently establishing the mapping relationship between world coordinates and pixel coordinates.

[0009] By matching the acquisition timestamps of the spatial coordinate data with the frame timestamps of the synchronized video data, a set of calibration point pairs containing world coordinates and pixel coordinates is established. Time matching is a crucial step in ensuring the correct correspondence between spatial coordinates and image positions. Through a precise time synchronization strategy, each spatial coordinate point is associated with the pixel position in the corresponding video frame, forming a calibration point pair. These calibration point pairs contain three-dimensional position information in the world coordinate system and two-dimensional pixel positions on the image plane, serving as the fundamental data for calculating camera calibration parameters.

[0010] Based on the geometric distribution characteristics of roads within the monitoring area, the calibration point pair set is divided into regions, obtaining multiple calibration sub-regions and their corresponding sub-region calibration point pairs. The region division fully considers the spatial distribution characteristics of roads in highway scenarios, such as the geometric characteristics of different areas like straight sections, curved sections, and ramp areas. Through a scientific region division strategy, the calibration point pairs are divided into multiple sub-regions, with calibration points within each sub-region exhibiting relatively consistent spatial distribution characteristics, which helps improve the calibration accuracy and adaptability of local areas.

[0011] An adaptive homography matrix is ​​calculated for each calibration sub-region's calibration point pairs to obtain the pixel-world coordinate mapping relationship for each calibration sub-region. Cross-validation is used to evaluate the generalization ability of the mapping relationship. Adaptive homography matrix calculation is the core technology of this method. By calculating dedicated mapping matrices for different sub-regions, accurate local calibration is achieved, overcoming the limitation of a single global mapping model being unable to adapt to complex scenarios. The cross-validation mechanism ensures that the calculated mapping relationship has good generalization ability, avoids overfitting, and improves the reliability of the calibration results in practical applications.

[0012] Based on the feature matching relationship between the currently acquired image and the reference image, the drift amount of the camera calibration parameters is obtained; the calibration validity is determined based on the drift amount. After calibration, the drift amount of the calibration parameters can be obtained as needed to evaluate the current calibration status.

[0013] In this embodiment of the invention, acquiring spatial coordinate data and synchronized video data of the monitoring area in a highway scenario includes: Within the permitted work area on highways, static point mapping is performed using a handheld RTK device to collect static spatial coordinate data of ground feature points. Static point mapping primarily targets fixed feature points such as road marking intersections, road signs, and median endpoints. These points are stable, easily identifiable, and ideal for establishing calibration benchmarks. During the data collection process, operators must remain within a safe permitted work area, typically an emergency lane or a temporarily closed lane, to ensure operational safety. The static spatial coordinate data records the precise three-dimensional position information of each feature point, including longitude, latitude, elevation, and a UTC timestamp at the time of data collection.

[0014] In highway scenarios, continuous trajectory acquisition is performed within the driving lane area using a vehicle-mounted RTK device fixed to the roof of a work vehicle to obtain dynamic spatial coordinate data. Dynamic trajectory acquisition compensates for the inability of static acquisition to cover the core area of ​​the driving lane. The work vehicle carrying the RTK device travels along a preset path, continuously collecting the spatial coordinates of the vehicle's trajectory points. The vehicle-mounted RTK device is typically installed in the center of the vehicle roof and secured with a strong bracket to ensure positional stability during acquisition. The dynamic spatial coordinate data contains densely distributed three-dimensional position points along the vehicle's travel path, providing rich calibration data for inaccessible core areas of the highway.

[0015] During the acquisition process, the camera records video synchronously, embedding SEI (Search Engine Information) supplementary enhancement information with a UTC millisecond-level timestamp into the video stream. Synchronous recording is a key measure to ensure accurate correspondence between spatial coordinates and image data. The camera continuously records video throughout the entire data acquisition process, embedding precise time information into the video stream. SEI supplementary enhancement information is a data carrier supported by modern video coding standards such as H.264 / H.265, which can carry additional metadata without affecting normal video playback. This method embeds a UTC millisecond-level timestamp into each video frame, achieving high-precision time synchronization between video frames and spatial coordinates.

[0016] Spatial coordinate data includes static spatial coordinate data and dynamic spatial coordinate data. These two types of data complement each other. Static data provides stable reference points, while dynamic data provides broad coverage. Together, they constitute a complete spatial coordinate dataset, laying the foundation for comprehensive and accurate camera calibration.

[0017] In this embodiment of the invention, establishing a set of calibration point pairs containing world coordinates and pixel coordinates includes: SEI supplementary enhancement information is parsed from video frames of synchronized video data to obtain the UTC millisecond-level timestamp for each video frame. The parsing process is implemented through the extended functionality of the video decoder or a dedicated SEI parsing module, extracting and decoding the SEI information from the video bitstream to obtain the embedded timestamp data. These timestamps are accurate to the millisecond level, providing a high-precision reference for subsequent time matching.

[0018] For each static point in the static spatial coordinate data, the corresponding video frame is located based on its acquisition timestamp. First, the RTK phase center coordinates are vertically projected and corrected according to the probe height. Then, the pixel position of the grounding point at the bottom of the probe is located in the video frame using pose recognition or manual annotation, establishing a static point pair. The positioning process uses the nearest timestamp matching principle to find the video frame closest to the static point acquisition time. Since the operator holds the RTK probe stationary during static point acquisition, the probe position can be clearly identified in the video frame. Annotation is performed manually or semi-automatically to accurately locate the pixel position of the probe's bottom contacting the ground, forming a static calibration point pair with the corresponding world coordinates.

[0019] For each trajectory point in the dynamic spatial coordinate data, a nearest neighbor matching strategy is used to find video frames with a time difference less than a preset time threshold. The pixel position of the grounding point of the vehicle where the RTK is located is marked in the video frame, establishing a dynamic calibration point pair. The nearest neighbor matching strategy ensures the precise correspondence between the dynamic trajectory point and the video frame. The preset time threshold is typically set to half the video frame interval (e.g., 20ms for a 25fps video). Matches exceeding the threshold are considered unreliable and discarded. In the matched video frames, the pixel position of the grounding point of the vehicle where the RTK is located is marked using computer vision algorithms or manual assistance, forming a dynamic calibration point pair with the corresponding world coordinates.

[0020] Static and dynamic calibration point pairs are merged to obtain a calibration point pair set. The merging process involves unified management and quality control of both types of calibration point pairs, eliminating obviously abnormal point pairs to form a complete calibration point pair set. This set contains rich calibration data from different parts of the covered monitoring area, providing comprehensive input for subsequent area division and homography matrix calculation.

[0021] In this embodiment of the invention, the detailed implementation steps for dividing the calibration point pair set into regions based on the geometric distribution characteristics of roads within the monitoring area, and obtaining multiple calibration sub-regions and their corresponding sub-region calibration point pairs, include: Based on the spatial distribution characteristics of the monitoring area and the camera's field of view, a region division strategy is determined. This strategy is crucial for accurate calibration and requires comprehensive consideration of factors such as road geometry, camera field of view angle, and changes in line-of-sight. For highway scenarios, a combined division method based on line-of-sight depth and lateral lane distribution is adopted, dividing the monitoring area into near, medium, and far-distance zones, and forming a grid-like region division structure based on lateral position. This division method conforms to the camera imaging principles and road spatial distribution characteristics, effectively addressing imaging differences in different areas.

[0022] Calculate the spatial distribution characteristic parameters of each point pair in the calibration point pair set.

[0023] The spatial distribution characteristic parameters of the point pairs are calculated to initially divide the calibration sub-regions. Before dividing, it is ensured that the calibration point pairs have undergone height vector compensation, i.e., the world coordinates have been vertically projected from the antenna phase center to the ground. The spatial characteristic parameters of each ground reference point in the point pair set relative to the camera are calculated. Based on the pre-defined polygonal regions of interest (ROIs) in the image plane, the inclusion relationship algorithm of points within polygons in computational geometry is used to directly assign the calibration point pairs to the corresponding calibration sub-regions. This process fully integrates the ground grounding point pixel coordinates extracted by YOLO-pose to ensure that each calibration point is accurately assigned to the imaging sub-block corresponding to its physical location.

[0024] Optimize the sub-region boundaries to ensure the uniform distribution and balanced number of calibration points within each sub-region. Adjust the initial division results to improve calibration quality. The optimization process considers three key factors: first, ensuring each sub-region contains a sufficient number of calibration points (usually no less than 20); second, ensuring the uniform distribution of points within each sub-region to avoid local clustering; and finally, adjusting the sub-region boundaries to conform as closely as possible to the road's geometry, such as dividing along lane lines or median strips. Through these optimization measures, a reasonable calibration sub-region division scheme is finally determined.

[0025] The calibration point pair sets are assigned to corresponding sub-regions according to the optimized sub-region partitioning results, resulting in multiple sub-region calibration point pair sets. The assignment process is the final step in region partitioning, explicitly assigning each calibration point pair to a specific calibration sub-region based on its spatial distribution characteristics. Points located near region boundaries may be assigned to two adjacent sub-regions simultaneously to enhance calibration continuity in the boundary regions. The calibration point pair sets for each sub-region are stored and managed separately, providing input data for subsequent homography matrix calculations.

[0026] In this embodiment of the invention, adaptive homography matrix calculation is performed on the sub-region calibration point pairs of each calibration sub-region, including: The original geographic coordinates (such as WGS84 latitude and longitude coordinates) of the sub-region calibration point pairs are transformed to a local Cartesian coordinate system with the camera position as the origin. The specific transformation process is as follows: A geographic coordinate projection algorithm is used to convert the ground reference points and camera positions into planar coordinates. Then, the origin is moved to the camera's location through coordinate translation to obtain local transition coordinates. Furthermore, a rotation transformation matrix is ​​constructed using the extracted road heading angle to rotate and align the local coordinates, ensuring that the coordinate axes are completely consistent with the actual vehicle driving direction or camera orientation. After the transformation, the world coordinates of all calibration points are decoupled into lateral and longitudinal displacements relative to the camera position, thereby eliminating the interference of sensor height factors on the planar mapping calculation and providing a data foundation for subsequent independent evaluation of projection errors in each dimension.

[0027] The basic reprojection threshold is determined based on the number of calibration point pairs in the sub-region. The basic reprojection threshold is then adjusted by a factor of magnitude based on the maximum observation distance of the corresponding calibration sub-region to obtain an adaptive reprojection threshold. The reprojection threshold is a key parameter in the RANSAC algorithm for distinguishing between interior and exterior points, directly affecting the algorithm's robustness and accuracy. The basic reprojection threshold is dynamically assigned in a stepwise manner based on different preset intervals containing the number of calibration points.

[0028] The RANSAC algorithm is employed to solve for the homography matrix of the transformed sub-region calibration point pairs based on an adaptive reprojection threshold, obtaining the pixel-world coordinate mapping relationship. The RANSAC algorithm is then used iteratively to calculate the homography matrix of the transformed ground calibration point pairs based on the aforementioned adaptive reprojection threshold to eliminate interference from noise-infected or occluded anomalous point pairs. This algorithm ultimately yields the optimal 3×3 homography transformation matrix, which accurately describes the mapping relationship between the ground reference world coordinates (lateral / longitudinal displacement) and the image ground point pixel coordinates extracted using YOLO-pose. This obtained matrix serves as the core result of camera calibration, used subsequently for real-time reconstruction of image pixels into vehicle trajectories and positions in real-world geographic space in high-speed scenes.

[0029] In this embodiment of the invention, the evaluation of the generalization ability of the mapping relationship based on cross-validation includes: Each labeled sub-region's labeled point pairs are randomly divided into training and test sets according to a preset ratio. Cross-validation is a classic method for evaluating a model's generalization ability, verifying its stability and applicability by testing the model's performance on different subsets of data. In this method, the preset ratio is typically 7:3 or 8:2, meaning 70%-80% of the labeled point pairs are used for training (to calculate the homography matrix), and the remainder for testing (to verify matrix accuracy). The partitioning process employs a hierarchical random sampling strategy to ensure that the training and test sets have similar spatial distribution characteristics, avoiding the impact of sampling bias on the validation results.

[0030] The homography matrix is ​​calculated using only the training set. This step is similar to the RANSAC algorithm described above, but only the partitioned training set data is used for calculation. The training set typically contains a sufficient number of calibration point pairs to support stable and reliable homography matrix calculation. The result is a 3×3 transformation matrix representing the pixel-to-world coordinate mapping obtained based on the training data.

[0031] Calculate the average reprojection error of the homography matrix on both the training and test sets. The average reprojection error is a direct indicator of calibration quality, reflecting the accuracy of the mapping relationship. The calculation process includes: projecting world coordinates to pixel coordinates (or backprojecting) using the homography matrix; calculating the Euclidean distance between the projected result and the actual labeled location; and calculating the average distance error for all points. This calculation is performed on both the training and test sets, yielding two error values, reflecting the model's performance on known and unknown data, respectively.

[0032] The system compares the average reprojection error between the test set and the training set. If the error on the test set is significantly higher than that on the training set (e.g., exceeding a preset ratio threshold), the calibration model for that sub-region is deemed to have an overfitting risk, meaning the model overfits a specific sampling trajectory and lacks generalization ability across the entire road segment. In this case, the system will trigger automatic optimization logic, re-dividing the sub-region or adjusting the RANSAC sampling weights to ensure that the final calibration parameters can support accurate vehicle positioning and speed measurement across the entire field of view.

[0033] In this embodiment of the invention, the drift amount of the camera calibration parameters is obtained based on the feature matching relationship between the currently acquired image and the reference image, including: Several frames are extracted from the current video recording as the currently acquired image for image feature extraction and matching. The image to be inspected is extracted from the current video frame, and the SIFT algorithm is used to extract key feature points. These points are then matched with the reference image saved during calibration to obtain initial matching point pairs that reflect the correspondence of scene structures.

[0034] Abnormal matching points are filtered based on the camera's preset pixel threshold. Matching points are verified using preset pixel thresholds (such as the difference between horizontal and vertical coordinates), and abnormal matching points with excessive deviations are removed by combining region filtering rules (such as excluding non-road interference areas at the edge of the image) to improve the accuracy of displacement estimation.

[0035] Quantitatively calculate the pixel-level physical parameter offset of the camera. The RANSAC algorithm is used to robustly estimate the homography transformation matrix from the current image to the reference image. By calculating the average Euclidean distance between the transformed feature point set and the reference point set, the drift in pixels is obtained, intuitively reflecting the physical parameter offset of the camera caused by environmental factors (such as vibration and temperature difference).

[0036] The system uses a preset absolute pixel drift threshold to provide tiered warnings for camera drift, assisting in camera calibration and maintenance. A drift exceeding a first drift threshold is considered a significant drift; a drift between a second and first drift threshold is considered a minor drift. The system presets an absolute pixel drift threshold; in a preferred embodiment, this threshold is configured differently based on the field of view and installation location of different cameras, typically set within the range of 5 to 10 pixels, representing a drift level that requires attention but does not currently affect normal system operation.

[0037] This invention achieves automated camera calibration in highway scenarios through high-precision spatial coordinate data acquisition, accurate time matching, scientific region division, adaptive homography matrix calculation, cross-validation evaluation, and real-time drift monitoring. The method of this invention can adapt to the special needs of highway scenarios, solving the accuracy and stability problems of traditional calibration methods under conditions such as long line-of-sight, large angle variations, and complex environments. It provides a reliable coordinate mapping foundation for video analysis and intelligent monitoring in highway scenarios.

[0038] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0039] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. An automated camera calibration method for highway scenarios, characterized in that, include: Acquire spatial coordinate data and synchronous video data of the monitoring area in a highway scenario; The acquisition timestamp of the spatial coordinate data is matched with the frame timestamp in the synchronized video data, and the target pixel coordinates are extracted using a target detection and tracking algorithm to establish a set of calibration point pairs containing world coordinates and pixel coordinates. The calibration point pair set is divided into regions based on the geometric distribution characteristics of roads within the monitoring area to obtain multiple calibration sub-regions and their corresponding sub-region calibration point pairs; For each calibration sub-region, an adaptive homography matrix is ​​calculated to obtain the pixel-world coordinate mapping relationship of each calibration sub-region. The generalization ability of the mapping relationship is evaluated based on cross-validation; The adaptive homography matrix calculation for the calibration point pairs of each calibration sub-region includes: Transform the original geographic coordinates of the sub-region calibration point pairs to a local Cartesian coordinate system with the camera position as the origin; The basic reprojection threshold is determined based on the number of calibration point pairs in the sub-region; the basic reprojection threshold is then corrected by a magnification based on the maximum observation distance of the corresponding calibration sub-region to obtain an adaptive reprojection threshold. The RANSAC algorithm is used to solve the homography matrix of the transformed sub-region calibration point pairs based on the adaptive reprojection threshold, thereby obtaining the pixel-world coordinate mapping relationship; Based on the feature matching relationship between the currently acquired image and the reference image, the drift amount of the camera calibration parameters is obtained; the calibration validity is determined based on the drift amount.

2. The automated camera calibration method for highway scenarios according to claim 1, characterized in that, The acquisition of spatial coordinate data and synchronous video data of the monitoring area in the highway scenario includes: Within the permitted work area in a highway scenario, static spatial coordinate data is obtained by manually using a handheld RTK device to statically mark ground feature points. In the driving lane area of ​​a highway scenario, continuous trajectory acquisition is performed using an on-board RTK device fixed to the top of the work vehicle to obtain dynamic spatial coordinate data; During the acquisition process, synchronous video recording is performed via camera, and SEI supplementary enhancement information with UTC millisecond-level timestamps is embedded in the video stream of the synchronous video data. The spatial coordinate data includes the static spatial coordinate data and the dynamic spatial coordinate data.

3. The automated camera calibration method for highway scenarios according to claim 2, characterized in that, The establishment of a set of calibration point pairs containing world coordinates and pixel coordinates includes: SEI supplementary enhancement information is parsed from the video frames of the synchronized video data to obtain the UTC millisecond-level timestamp of each video frame; For each static point in the static spatial coordinate data, locate the corresponding video frame according to its acquisition timestamp, obtain the height parameter of the RTK measuring rod, and perform height vector compensation projection on the static spatial coordinate data to obtain the ground reference world coordinates. Mark the pixel position of the bottom of the RTK measuring rod in the video frame to establish a static calibration point pair. For each trajectory point in the dynamic spatial coordinate data, the nearest neighbor matching strategy is used to find video frames with a time difference less than a preset time threshold; the YOLO target detection and pose recognition algorithm is used to extract the grounding point pixel coordinates of the vehicle target in the video frame; and the ByteTrack algorithm is introduced to track the target across frames and assign a unique ID to establish dynamic calibration point pairs. The static calibration point pairs and the dynamic calibration point pairs are merged to obtain the calibration point pair set.

4. The automated camera calibration method for highway scenarios according to claim 1, characterized in that, The evaluation of the generalization ability of the mapping relationship based on cross-validation includes: The sub-region calibration point pairs of each calibration sub-region are randomly divided into training set and test set according to a preset ratio; The homography matrix is ​​computed using only the training set; Calculate the average reprojection error of the homography matrix on the training set and the test set, respectively; When the difference between the average reprojection error on the test set and the average reprojection error on the training set exceeds a preset error threshold, it is determined that the mapping relationship is overfitted.

5. The automated camera calibration method for highway scenarios according to claim 1, characterized in that, The step of obtaining the drift amount of camera calibration parameters based on the feature matching relationship between the currently acquired image and the reference image includes: Randomly select several frames from the current video recording as the currently captured image; SIFT features are extracted from the currently acquired image and the reference image saved during calibration, and the extracted SIFT features are matched to obtain matching point pairs. Calculate the homography transformation matrix from the currently acquired image to the reference image based on the matching point pairs; The feature point set is transformed using the calculated homography matrix, and the average Euclidean distance between the transformed point set and the reference point set is calculated. This average Euclidean distance is used as the drift amount. When the drift amount is greater than the first drift threshold, it is determined to be a large drift; when the drift amount is between the second drift threshold and the first drift threshold, it is determined to be a slight drift.

Images