A method for measuring column dimensions and locating corner points based on UAV technology and computer vision, and a method for column positioning.

Abstract

This invention belongs to the field of automatic measurement in civil engineering, and relates to a method for column dimension measurement and corner point localization based on UAV technology and computer vision. The method includes: acquiring orthophotos of the engineering site; coarsely locating the columns in the orthophotos based on a target detection model and marking coarse positioning boxes; segmenting the columns according to the coarse positioning box information to obtain single-column datasets, extracting the column inner contours from the single-column datasets, and calculating the column dimensions and corner point coordinates; fusing the coarse positioning box coordinates and column corner point coordinates to obtain the image coordinates of the column in the orthophotos; and performing an affine transformation to obtain the column world coordinates. This invention completes the dimension measurement and corner point localization of columns at the engineering site through a technical route of "coarse positioning → contour extraction → joint precise positioning," ultimately achieving column localization. It systematically solves the problems of insufficient positioning accuracy of rotated columns, difficulty in achieving sub-pixel accuracy of contours, and low measurement efficiency at the engineering site in traditional methods.

Description

Technical Field

[0001] This invention belongs to the field of automatic measurement in civil engineering, and more specifically, relates to a method for measuring column dimensions and locating corner points based on UAV technology and computer vision, as well as a method for column positioning. Background Technology

[0002] Columns are the main load-bearing components of a building, transferring the load from the superstructure to the foundation and subsoil, ensuring the overall stability of the structure. Therefore, their geometric and spatial positioning accuracy directly affects construction quality and structural safety. Traditional column positioning relies on manual labor, using contact measurement methods such as total stations and RTK-GPS, which has significant limitations: setting up instruments in complex terrain (water areas, deep foundation pits) is time-consuming, and for large sites with many columns, manual measurement is time-consuming and labor-intensive. In addition, manual measurement in complex scenarios may pose certain dangers.

[0003] In recent years, with the rapid development of UAV technology and computer vision, UAVs, with their advantages of high flexibility, high operational efficiency, high safety, and low cost, have been widely used in various industries for intelligent decision-making by integrating computer vision image processing, feature extraction, and pattern recognition technologies. This has also provided new ideas for the field of civil engineering inspection and monitoring. However, there is currently no research on column positioning in complex scenarios. Directly using existing technologies faces two challenges: First, insufficient feature extraction for rotating targets: traditional horizontal detection boxes suffer from excessive errors in column center point positioning due to boundary redundancy. Second, interference from complex backgrounds: factors such as sunlight shadows and temporary support structures increase the false detection rate of edge extraction algorithms. Current research focuses on general target detection, such as people and vehicles, while there are no good solutions for the special needs of column positioning (such as column boundary constraints). For example, while the YOLO series algorithms achieve 95% mAP in horizontal target detection, their positioning accuracy for rotating columns does not meet engineering requirements. Deep learning edge detection algorithms such as EdgeNAT can initially extract edges, but they struggle to extract contours and achieve sub-pixel accuracy, failing to meet engineering needs. To address these issues, we developed methods for measuring column dimensions and locating corner points, which improves measurement efficiency on-site and is of great significance for achieving automation and intelligence in measurement within the civil engineering field. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a column dimension measurement and corner point positioning method based on UAV technology and computer vision, as well as a column positioning method, the purpose of which is to solve the technical problem of improving the measurement efficiency on the engineering site.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for measuring column dimensions and locating corner points based on UAV technology and computer vision is provided, comprising:

[0006] S1: Acquire orthophotos of the construction site;

[0007] S2: Based on the target detection model, coarsely locate the columns in the orthophoto of the engineering site and mark the coarse location boxes;

[0008] S3: Cut the single-column dataset based on the coarse positioning box information, extract the inner contour of the column from the single-column dataset, and calculate the column size and corner coordinates.

[0009] Further, step S1 includes:

[0010] Collect multi-lens image data from the drone, import the images with POS information into 3D reality modeling software, and perform the first aerial triangulation calculation;

[0011] Import control points and perform a second aerial triangulation calculation;

[0012] Align the image data with the actual geographic coordinate system and perform geometric adjustment to reduce cumulative errors;

[0013] A TIN surface model is constructed using point cloud data, and after digital differential correction, an orthophoto of the engineering site is output.

[0014] Furthermore, the object detection model in step S2 includes Faster R-CNN, R-FCN, SSD, or YOLO models.

[0015] Furthermore, the object detection model in step S2 is YOLO11-OBB, and the offline training method for this model is as follows:

[0016] Acquire multi-camera image data from a drone, and perform multi-angle rotation enhancement processing on each image from the downward-looking camera with different angle sequences to meet the training requirements for rotating target detection; label and convert the images to obtain dataset images in YOLO11-OBB format, from which a training set is divided; the labeled categories include pillars, concrete mixer trucks, excavators and / or trucks;

[0017] Perform data augmentation operations on the labeled training set;

[0018] The enhanced training set images are input into the YOLO11-OBB network, initial weight parameters are set, and a rotating cylinder target detection model is trained to perform coarse cylinder localization as the target detection model.

[0019] Furthermore, step S2 also includes lossless rotation of the coarse localization results of the target detection model, including:

[0020] The orthophoto detection result includes the detected image and information labels containing the categories and coordinates of the image. The columns in the image are filtered out by category number to obtain their detection box coordinate information.

[0021] Based on the coordinate information of the detection box, the coordinates of the center point of the detection box, the rotation angle of the detection box relative to the x-direction of the image coordinate system, and the length and width of the box are calculated.

[0022] The rotation angles of each detection box are arranged according to multiple intervals. The interval containing the most rotation angles is selected, and the average rotation angle in that interval is calculated to determine the overall rotation angle of the image.

[0023] Using the image center as the rotation center, construct a lossless rotation matrix for the image, so that after rotation, all column detection boxes are parallel to the column base edge and at the same time, other image information is not lost;

[0024] To ensure that the rotated image is lossless, the original image area is expanded to complete the rotation of the image and the center point coordinates of the column detection box.

[0025] Furthermore, the column detection results are optimized after lossless rotation, including:

[0026] Calculate the maximum value of the length and width of all columns, and use it as the length and width of all column detection boxes to ensure that all detection boxes can contain the entire column area, thereby achieving coarse column positioning.

[0027] The image is cropped based on the location of the bar detection boxes to obtain a single bar image set, and the coordinates of the top left corner of each detection box are calculated.

[0028] Change the coordinate format of the column detection box to include the column number, the center coordinates of the rotated column, the coordinates of the top left corner of the rotated column detection box, and the length and width of the column detection box.

[0029] Further, step S3 includes the following steps:

[0030] The cropped single-column dataset was labeled using an image segmentation tool and converted into a binary image as a labeling file, resulting in a labeled single-column dataset image.

[0031] After enhancing the labeled single-column dataset images, edge detection is performed, and the edge detection results of each column are extracted.

[0032] Morphological erosion is performed on the edge pixels extracted from the edge detection results to remove noise. Then, the inner edge of the cylinder is fitted to obtain the coordinates of the inner edge and the four corner points.

[0033] The sub-pixel coordinates of the four corner points of each pillar are calculated by using the Gaussian approximation method.

[0034] The cylinder size is calculated based on the obtained sub-pixel coordinates of the corner points.

[0035] According to another aspect of the present invention, a column positioning method based on UAV technology and computer vision is provided, including the column size measurement and corner point positioning method based on UAV technology and computer vision as described in any of the preceding claims, and further including the following steps:

[0036] S4: Merge the coordinates of the coarse positioning frame and the coordinates of the column corner points to obtain the image coordinates of the column in the orthophoto;

[0037] S5: Perform an affine transformation to obtain the cylindrical world coordinates.

[0038] Further, the image coordinates of the cylinder in the orthophoto in step S4 include:

[0039] Using the column sub-pixel corner coordinates containing column contour information obtained in step S3, the coarse positioning coordinates of the column obtained in step S2 are transformed, and the transformation result is subjected to inverse rotation matrix transformation to obtain the image coordinates of the column in the original orthophoto.

[0040] Further, step S5 includes:

[0041] Based on the affine parameters obtained when calculating the column dimensions in step S3, an affine matrix is ​​constructed to solve for the world coordinates of the four corner points of the column, as well as the column's outline information, thus obtaining the column's world coordinates.

[0042] In summary, the technical solutions conceived in this invention, compared with the prior art, can achieve the following beneficial effects:

[0043] 1. This invention proposes a column dimension measurement and corner point positioning method based on UAV technology and computer vision. Through coarse positioning and cutting extraction operations, it greatly improves the efficiency of on-site engineering measurement and is of great significance for realizing the automation and intelligence of measurement in the field of civil engineering.

[0044] 2. Preferably, during the generation of orthophotos of the engineering site, two aerial triangulation calculations and geometric adjustment are performed to reduce cumulative errors. Combined with digital differential correction of the point cloud model, the accuracy of the original orthophoto influence of the engineering site can be significantly improved, thereby improving the quality and accuracy of the final positioning result from the source.

[0045] 3. Preferably, during the training process of the target detection model, the labeled categories include pillars, concrete mixer trucks, excavators and / or trucks, and data augmentation is performed. This can avoid overfitting during the training process, improve the model's generalization ability, and further improve the quality and accuracy of the initial detection results of pillar targets.

[0046] 4. Preferably, while most cylindrical targets can be coarsely located using the target detection model, there are still some cases where the features of rotating cylindrical targets are insufficiently extracted. By performing operations such as center detection, interval averaging, constructing a lossless rotation matrix, and region expansion, the comprehensiveness and completeness of the cylindrical detection information can be greatly improved.

[0047] 5. Preferably, the "maximum box - fitting box" method is used to fit the inner edge of the column, which can further improve the accuracy of edge extraction.

[0048] 6. Preferably, the sub-pixel coordinate extraction and cylinder size calculation based on the Gaussian approximation method can overcome the problem of reduced engineering accuracy caused by the limitation of the pixel resolution of the image itself.

[0049] 7. Preferably, combining coarse positioning and contour information for precise column positioning and then performing affine transformation can improve the computational efficiency of precise positioning and achieve the effect of high-precision and high-efficiency positioning of the world coordinates of the column.

[0050] 8. This invention proposes a column size measurement and corner point positioning method based on UAV technology and computer vision. Through the technical route of "coarse positioning → contour extraction → joint precise positioning", the column size measurement and corner point positioning of the engineering site are completed, and the column positioning is finally realized. This systematically solves the problems of insufficient positioning accuracy of rotating columns, difficulty in achieving sub-pixel accuracy of contours, and low measurement efficiency in engineering sites in traditional methods. Attached Figure Description

[0051] Figure 1 This is a general flowchart of a preferred embodiment of the present invention;

[0052] Figure 2 These are orthophotos of the engineering site according to a preferred embodiment of the present invention;

[0053] Figure 3 This is the target detection result of a preferred embodiment of the present invention;

[0054] Figure 4 This is the coarse positioning box marking result of a preferred embodiment of the present invention;

[0055] Figure 5 'a' is a partial example of the cropped single-column dataset of a preferred embodiment of the present invention;

[0056] Figure 5 b is a partial example of the column edge detection results in a preferred embodiment of the present invention;

[0057] Figure 6 This is a schematic diagram of the inner edge fitting steps in a preferred embodiment of the present invention;

[0058] Figure 7a~f are six examples of fitting box results in a preferred embodiment of the present invention, wherein the green box is the largest box, the red box is the outer edge box of the column, and the blue box is the inner edge box of the column;

[0059] Figure 8 This is an example of the final fitted box of the inner edges of the two pillars in a preferred embodiment of the present invention, where a corresponds to... Figure 7 a and b correspond to Figure 7 b;

[0060] Figure 9 This is an example of a subpixel coordinate fitting box for the inner edges of two pillars in a preferred embodiment of the present invention, where 'a' corresponds to... Figure 8 a and b correspond to Figure 8 b;

[0061] Figure 10 yes Figure 9 Column 1 (i.e. Figure 9 The diagram below shows the solution for the subpixel coordinates of the top left corner point (example a). The top (left) image is the new image (with the top left corner point) after superimposing the grayscale image of the original image of column 1 and its edge image; the top (middle) image is the initialization graphic of the two-dimensional Gaussian function; the top (right) image is the fitted two-dimensional Gaussian function; and the bottom image shows the original coordinates and subpixel coordinates of the top left corner point of column 1.

[0062] Figure 11 This is an example of a precise positioning frame according to a preferred embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0064] A method for measuring column dimensions and locating corner points based on UAV technology and computer vision is proposed. The technical route includes "coarse positioning → contour extraction → joint precise positioning", and the specific method is as follows.

[0065] S1 Orthophoto Generation Based on Control Points

[0066] Five-lens image data from a UAV is collected. Images with POS information are imported into ContextCapture 3D reality modeling software for the first aerial triangulation calculation. Human-made control points are then imported for a second aerial triangulation calculation to align the image data with the actual geographic coordinate system. Geometric adjustment is performed to reduce cumulative errors. A TIN surface model is constructed using point cloud data generated by a high-precision matching algorithm. After digital differential correction, an orthophoto of the engineering site is output (e.g., ...). Figure 2 ).

[0067] S2 Coarse Pillar Localization Based on Orthophoto Imagery Using Target Detection Network

[0068] Common object detection networks include Faster R-CNN, R-FCN, SSD, and YOLO models. Among them, the YOLO series of models has been continuously iterated and updated, and the detection accuracy has been continuously improved. This embodiment uses the latest YOLO11-OBB as an example for illustration.

[0069] YOLO11-OBB Network Training and Pillar Detection:

[0070] Five-lens image data from a UAV were acquired. Each image from the downward-facing lens was subjected to multi-angle rotation enhancement processing in an angle sequence of (5°, 10°, ..., 45°) to meet the training requirements for rotating target detection. The RoLabelImg data annotation tool was used for annotation, and the format was converted to obtain a dataset image in YOLO11-OBB format, which was further divided into training, validation, and test sets. To avoid overfitting, the labeled categories included columns, concrete mixer trucks, excavators, and trucks. Data augmentation operations were performed on the labeled training set, including random cropping, flipping, scaling, geometric transformations, and color transformations adjusting brightness, exposure, and saturation to improve the model's generalization ability. The enhanced training set images were input into the YOLO11-OBB network, initial weight parameters were set, and a rotating column target detection model was trained. Orthophotos of the engineering site were input into the model for detection to obtain column target detection results (e.g., ...). Figure 3 ).

[0071] Lossless rotation of orthophotos and optimization of cylinder detection results:

[0072] Using a trained YOLO11-OBB model for column detection, most columns can be coarsely located. However, some columns still suffer from insufficient feature extraction, such as: incomplete column information (the detection box fails to include the entire column area); and boundary redundancy (excessive error in column center point localization). In the coarse column localization stage, the latter can be temporarily ignored, while the former requires result optimization to achieve coarse localization of all columns. The method is as follows:

[0073] Orthophoto detection results include the detected image and information labels containing the categories and coordinates of the images. The columns in the image are filtered by category number to obtain their bounding box coordinates, in the following format:

[0074]

[0075] in For column number, These are the coordinates of the four corner points. , The coordinates of the center point of the detection frame are calculated based on the coordinate information of the four corner points. The rotation angle of the detection box relative to the x-axis of the image coordinate system. and frame length and frame width :

[0076]

[0077]

[0078]

[0079]

[0080]

[0081] First, rotate the angle of each detection frame. according to Arrange the 45 intervals and select the one containing the most The interval, by calculating the rotation angle within it. The average value is used to determine the overall rotation angle of the image. :

[0082]

[0083] With image center Construct a lossless rotation matrix for the image, with the rotation center as the rotation center. This ensures that after rotation, all column detection boxes are parallel to the edges of the column bases without losing other image information. W and H are the length and width of the image. αi

[0084] Generally based on a certain point Rotation matrix with rotation center for:

[0085]

[0086] However, the rotated image exceeds the original image boundary, causing partial cropping and information loss. Therefore, to ensure the rotated image is lossless, the original image region is expanded as follows:

[0087]

[0088]

[0089] , These are the enlarged length and width. After enlarging the original region, the rotation matrix... Become :

[0090]

[0091] To facilitate the subsequent calculation of the inverse of the rotation matrix, the matrix is ​​converted to homogeneous form, and the rotation center is set to the image center. The lossless rotation matrix obtained by this method is thus obtained. :

[0092]

[0093]

[0094] This allows us to obtain the center coordinates of the rotated cylinder. The coordinates are transformed into homogeneous form. Perform the calculation:

[0095]

[0096] At this step, the rotation of the image and the center point coordinates of the column detection boxes has been completed. However, the problem of incomplete column information still needs to be addressed. The coordinates of each column have already been calculated. ,calculate maximum value The length and width of all column detection boxes are used to ensure that each detection box encompasses the entire column region, thus achieving coarse column localization. Further, the image is cropped based on the location of the column detection boxes to obtain a single column image set. To facilitate the subsequent calculation of the precise column localization image coordinates, the coordinates of the top-left corner of the detection box are calculated here. :

[0097]

[0098] Change the coordinate format of the column detection box to: ,in For column numbering, retain necessary information and reduce formatting. The image of the coarsely located detection box is shown below. Figure 4 .

[0099] It should be noted that the above-mentioned lossless rotation of orthophotos and optimization method for cylinder detection results is a processing of target detection results, and therefore is not limited by the target detection model. Regardless of the target detection model used, as long as the orthophoto detection results contain the detection image and information labels containing the category and coordinates in the image, the above-mentioned method of the present invention can be applied.

[0100] S3 Single-cylinder contour information extraction based on edge detection and Gaussian approximation

[0101] Common edge detection algorithms and models include the Canny edge detection algorithm, the YOLO series models (which can be used for both object detection and edge detection), RetinaNet, EfficientDet, and EdgeNAT. Among them, EdgeNAT is a Transformer-based edge detection model that specifically uses the Dilated Neighborhood Attention Transformer (DiNAT) as its encoder, aiming to improve the efficiency and accuracy of edge detection. This example uses EdgeNAT to illustrate pillar edge detection.

[0102] EdgeNAT network training and pillar edge detection:

[0103] Use the LabelMe image segmentation tool (or other image annotation and segmentation tools such as LabelImg, RectLabel, OpenCV, Labelbox, CVAT, VIA-VGG Image Annotator, etc.) to process the cropped single-column dataset. Figure 5 The images in (a) are labeled, and their format is converted to generate corresponding binary images as label files, resulting in a BSDS format dataset. This dataset is further divided into training, validation, and test sets. Data augmentation operations are performed on the labeled training set, including scaling and color transformations such as brightness, exposure, and saturation adjustments, to improve the model's generalization ability. The augmented training set images are then input into the EdgeNAT network, initial weight parameters are set, and a column edge detection model is trained. The single-column images obtained after coarse column localization and cropping are input into the model for detection, yielding the edge detection results for each column. Figure 5 (b).

[0104] like Figures 6-8 As shown, since directly extracting the contours from the edge detection results of each column may result in inaccurate image edges due to noise, we first perform morphological erosion on the extracted edge pixels to remove noise. Then, we use the "maximum bounding box - fitted bounding box" method to fit the inner edges of the columns, that is, based on the size of the column detection box... Initial selection rectangle This is used to select the region containing the most edge pixels using a sliding bounding box, and then further fit the image inward within this region, setting certain judgment criteria: defining the overlap degree. =Number of pixels contained in the edge / Side length. Before moving each edge in the first step, determine whether movement is necessary to avoid incorrect fitting or the presence of individual noise points after the erosion operation. Movement is only stopped if the overlap after the movement is less than before the movement and the overlap before the movement is greater than 50%; otherwise, movement is performed. When each edge moves inward by one pixel, the overlap is calculated. If the overlap is greater than 75%, movement stops; otherwise, it continues until the overlap is greater than 75% and the fitting is complete. Each edge is fitted separately to ensure no cross-influence. The outer edge of the column is obtained through the above operations. The above operations are then continued, this time with the judgment criterion changed to: overlap less than 25%, further obtaining the column... i Inner edge and four corner points ( Coordinates of )

[0105]

[0106] Sub-pixel coordinate extraction and cylinder size calculation based on Gaussian approximation:

[0107] Due to the limitations of the image's pixel resolution, the obtained corner coordinates can only be integers, which does not meet engineering requirements. Since it is physically impossible to add more actual photosensitive units between adjacent pixels, a virtual pixel is introduced through an algorithm to improve the point positioning accuracy in the image.

[0108] In this invention, the Gaussian approximation method is used to calculate sub-pixel coordinates. First, the grayscale image and the edge detection image are superimposed at a 6:4 ratio to form a new image I (which preserves the details of the original image while ensuring that the calculated sub-pixel points are within a suitable range). Then, the new image I is equalized. Next, the Sobel operator is used to calculate the gradient magnitude of each pixel in the new image I. The area around the calculated corner pixels is divided into 3×3 regions. By fitting a two-dimensional Gaussian function, the location with the largest gradient magnitude is found. The horizontal and vertical coordinates at this location are the sub-pixel coordinates of the corner point. Figure 10 As shown, with corner points For example, the sub-pixel coordinates are solved as follows:

[0109] Horizontal and vertical gradients for each pixel:

[0110]

[0111] Therefore, the gradient magnitude of each pixel is:

[0112]

[0113] Accordingly, the corner points can be calculated. The gradient magnitude and coordinate information of 9 points within a 3×3 area (including the corner point) are represented as follows: ( x ,y , G By fitting a two-dimensional Gaussian function:

[0114]

[0115] The horizontal and vertical coordinates of the peak can be obtained. (i.e., sub-pixel coordinates) .

[0116] The calculation method for all other corner points is the same. The sub-pixel coordinates of the four corner points of each pillar are obtained through the Gaussian approximation method. .

[0117] The cylinder size is calculated based on the obtained sub-pixel coordinates of the corner points. When generating the orthophoto using ContextCapture, a .tfw file containing image information is also generated, which is then parsed into... ,in For image Orientation resolution, For image Orientation rotation coefficient, The rotation coefficient in the y-direction of the image. For image Orientation resolution, Centered at the top left pixel of the image coordinate, Centered at the top left pixel The coordinates and column dimensions are calculated as follows:

[0118]

[0119]

[0120] Save the coordinates of the column containing its contour information as follows:

[0121]

[0122] in For column numbers, retain necessary information, reduce formatting, and the fitting result is as follows: Figure 9 As shown.

[0123] S4 Combines coarse positioning and contour information for precise column positioning.

[0124] First, determine the coordinates of the cylinder in the orthophoto:

[0125] In section 3.2, coarse positioning of the column is achieved, with coordinates as follows: ,in For column numbers, These are the top left corner pixels of the detection box. Coordinates, combined with the cylinder sub-pixel corner coordinates containing contour information in section 3.3. The coordinates of the cylinder in the rotated image can then be determined. As follows:

[0126]

[0127] The calculations for the coordinates of other corner points are the same; the transformed cylindrical coordinates are as follows:

[0128]

[0129] Through lossless rotation matrix Inverse is obtained Thus, the image coordinates of the cylinder in the original orthophoto are calculated:

[0130]

[0131] Similarly, the coordinates of the column in the original orthophoto image, including its contour information, are obtained as follows:

[0132]

[0133] S5 uses affine transformation to find the world coordinates of the cylinder corner point:

[0134] The affine parameters were obtained from the .fw file when calculating the column dimensions. ,in For image Orientation resolution, For image Orientation rotation coefficient, The rotation coefficient in the y-direction of the image. For image Orientation resolution, Centered at the top left pixel of the image coordinate, Centered at the top left pixel Coordinates, construct an affine matrix to obtain the world coordinates of the corner points:

[0135]

[0136] The rest is the same; the final precise world coordinates and contour information of the cylinder corner points are as follows:

[0137]

[0138] in For column number, .

[0139] Precise positioning frame Figure 10 As shown, Figure 10The red coordinates represent the image coordinates of the four corner points of the column, the blue coordinates represent the world coordinates of the four corner points, and the yellow numbers represent the column dimensions, i.e., length and width, in meters.

[0140] Thus, this invention proposes a column dimension measurement and corner point positioning method based on UAV technology and computer vision. The method completes the column dimension measurement and corner point positioning on the engineering site through the technical route of "coarse positioning → contour extraction → joint precise positioning".

[0141] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for measuring column dimensions and locating corner points based on UAV technology and computer vision, characterized in that, include: S1: Acquire orthophotos of the construction site; S2: Based on the target detection model, coarsely locate the columns in the orthophoto of the engineering site and mark the coarse location boxes; Step S2 also includes lossless rotation of the coarse localization results of the target detection model, including: The orthophoto detection result includes the detected image and information labels containing the categories and coordinates of the image. The columns in the image are filtered out by category number to obtain their detection box coordinate information. Based on the coordinate information of the detection box, the coordinates of the center point of the detection box, the rotation angle of the detection box relative to the x-direction of the image coordinate system, and the length and width of the box are calculated. The rotation angles of each detection box are arranged according to multiple intervals. The interval containing the most rotation angles is selected, and the average rotation angle in that interval is calculated to determine the overall rotation angle of the image. Using the image center as the rotation center, construct a lossless rotation matrix for the image, so that after rotation, all column detection boxes are parallel to the column base edge and at the same time, other image information is not lost; To ensure that the rotated image is lossless, the original image area is enlarged to complete the rotation of the image and the coordinates of the center point of the column detection box; S3: Cut the single-column dataset based on the coarse positioning box information, extract the inner contour of the column from the single-column dataset, and calculate the column size and corner coordinates.

2. The method for measuring column dimensions and locating corner points based on UAV technology and computer vision as described in claim 1, characterized in that, Step S1 includes: Collect multi-lens image data from the drone, import the images with POS information into 3D reality modeling software, and perform the first aerial triangulation calculation; Import control points and perform a second aerial triangulation calculation; Align the image data with the actual geographic coordinate system and perform geometric adjustment to reduce cumulative errors; A TIN surface model is constructed using point cloud data, and after digital differential correction, an orthophoto of the engineering site is output.

3. The method for measuring column dimensions and locating corner points based on UAV technology and computer vision as described in claim 1, characterized in that, The object detection model in step S2 includes Faster R-CNN, R-FCN, SSD, or YOLO.

4. The method for measuring column dimensions and locating corner points based on UAV technology and computer vision as described in claim 3, characterized in that, The object detection model in step S2 is YOLO11-OBB, and the offline training method for this model is as follows: Acquire multi-camera image data from a drone, and perform multi-angle rotation enhancement processing on each image from the downward-looking camera with different angle sequences to meet the training requirements for rotating target detection; label and convert the images to obtain dataset images in YOLO11-OBB format, from which a training set is divided; the labeled categories include pillars, concrete mixer trucks, excavators and / or trucks; Perform data augmentation operations on the labeled training set; The enhanced training set images are input into the YOLO11-OBB network, initial weight parameters are set, and a rotating cylinder target detection model is trained to perform coarse cylinder localization as the target detection model.

5. A method for measuring column dimensions and locating corner points based on UAV technology and computer vision as described in any one of claims 1 to 4, characterized in that, After lossless rotation, the column detection results are optimized, including: Calculate the maximum value of the length and width of all columns, and use it as the length and width of all column detection boxes to ensure that all detection boxes can contain the entire column area, thereby achieving coarse column positioning. The image is cropped based on the location of the bar detection boxes to obtain a single bar image set, and the coordinates of the top left corner of each detection box are calculated. Change the coordinate format of the column detection box to include the column number, the center coordinates of the rotated column, the coordinates of the top left corner of the rotated column detection box, and the length and width of the column detection box.

6. The method for measuring column dimensions and locating corner points based on UAV technology and computer vision as described in claim 1, characterized in that, Step S3 includes the following steps: The cropped single-column dataset was labeled using an image segmentation tool and converted into a binary image as a labeling file, resulting in a labeled single-column dataset image. After enhancing the labeled single-column dataset images, edge detection is performed, and the edge detection results of each column are extracted. Morphological erosion is performed on the edge pixels extracted from the edge detection results to remove noise. Then, the inner edge of the cylinder is fitted to obtain the coordinates of the inner edge and the four corner points. The sub-pixel coordinates of the four corner points of each pillar are calculated by using the Gaussian approximation method. The cylinder size is calculated based on the obtained sub-pixel coordinates of the corner points.

7. A column localization method based on UAV technology and computer vision, characterized in that, The method for measuring column dimensions and locating corner points based on UAV technology and computer vision, as described in any one of claims 1 to 6, further includes the following steps: S4: Merge the coordinates of the coarse positioning frame and the coordinates of the column corner points to obtain the image coordinates of the column in the orthophoto; S5: Perform an affine transformation to obtain the cylindrical world coordinates.

8. The column localization method based on UAV technology and computer vision as described in claim 7, characterized in that, The image coordinates of the cylinder in the orthophoto in step S4 include: Using the column sub-pixel corner coordinates containing column contour information obtained in step S3, the coarse positioning coordinates of the column obtained in step S2 are transformed, and the transformation result is subjected to inverse rotation matrix transformation to obtain the image coordinates of the column in the original orthophoto.

9. A column localization method based on UAV technology and computer vision as described in claim 7, characterized in that, Step S5 includes: Based on the affine parameters obtained when calculating the column size in step S3, an affine matrix is ​​constructed to solve for the world coordinates of the four corner points of the column and the outline information of the column, thus obtaining the world coordinates of the column.

Images