A method for intelligently detecting deviation of stone column position

Abstract

The present application belongs to the technical field of gravel pile detection, and provides a gravel pile position deviation intelligent detection method, in order to solve the problems of low efficiency and large error of the existing detection method. The present application is based on a six-dimensional feature joint decision mechanism, simultaneously extracts six-dimensional features including blue, green, red, gray, contour area and circularity, and through multi-dimensional joint threshold decision, greatly reduces the false detection rate and the missed detection rate under the construction site environment such as complex illumination, ground debris and pile body shielding. The present application uses interactive measurement to obtain the average radius and inter-row spacing of the gravel pile, and automatically calculates the binary threshold value based on the Otsu algorithm. Compared with the global fixed threshold value, it can dynamically adapt to images under different illumination conditions, improve the robustness of foreground segmentation, and does not need manual repeated parameter adjustment. Different sites can directly reuse the process, solving the problem of relying on artificial experience and poor generalization ability of the traditional method.

Description

Technical Field

[0001] This invention belongs to the field of crushed stone pile detection technology, specifically relating to an intelligent detection method for crushed stone pile position deviation. Background Technology

[0002] Crushed stone piles are widely used in composite foundation treatment projects for special soil and rock types such as soft soil and miscellaneous fill. The accuracy of the pile placement directly determines the overall stress performance of the foundation. Therefore, pile position deviation is a core control item for the quality acceptance of engineering construction.

[0003] Currently, the technology for detecting pile position deviation of crushed stone piles mainly includes the following three application forms:

[0004] 1. Manual instrument inspection technology: This method relies on surveying equipment such as total stations and RTK-GPS to collect coordinates for each pile, and then manually compares them with the design data to complete the judgment. The accuracy of this method depends on the skill level of the operator, and human error is unavoidable. At the same time, the work of measuring each pile is arduous and the inspection efficiency is extremely low. It is impossible to achieve 100% full inspection for large construction sites, and the sampling inspection mode has blind spots in quality control.

[0005] 2. Traditional Aerial Photography-Assisted Inspection Technology: With the popularization of low-altitude aerial surveying technology, UAV orthophotos are gradually being applied to construction site acquisition, which can quickly reconstruct the overall site view. However, this technology is only used as an image acquisition method. Pile identification, pile core marking, and deviation calculation still need to be completed manually one by one. The degree of automation is low, and it does not fundamentally solve the problem of low inspection efficiency.

[0006] 3. Existing image recognition technology: Existing image recognition-based solutions mostly rely on single contour features to determine the pile body. They have weak anti-interference capabilities in complex lighting conditions, surface debris, pile body obstruction, and other scenarios on construction sites. They also have prominent problems of false detection and missed detection, which cannot meet the accuracy requirements of project acceptance.

[0007] In summary, existing crushed stone pile detection technologies suffer from drawbacks such as low efficiency, large human error, low automation, and poor adaptability to complex scenarios. Summary of the Invention

[0008] In order to solve or at least partially solve the above-mentioned technical problems, the present invention provides an intelligent detection method for pile position deviation of crushed stone piles.

[0009] To solve the technical problem, the technical solution adopted by this invention is as follows:

[0010] A smart detection method for pile position deviation of crushed stone piles includes:

[0011] S1: Image Acquisition and Construction Functional Module for Crushed Stone Piles: Acquire orthophotos of the crushed stone pile construction area; construct functional modules based on the OpenCV image processing library and Python language, including a sample annotation module, an intelligent recognition module, an image correction module, and a pile location detection module;

[0012] S2: Sample Labeling and Feature Rule File Generation: The average diameter and row-column spacing of the gravel piles in the orthophoto are calculated using the sample labeling module. Using the circular frame corresponding to the calculated average diameter as the labeling carrier, the gravel piles in the orthophoto are randomly and manually labeled. The number of labeled samples is no less than 60. Six types of feature parameters are extracted for each sample: including blue channel value B, green channel value G, red channel value R, gray value, contour area, and circularity. The binarization threshold is adaptively calculated based on the Otsu algorithm to generate the gravel pile feature rule file.

[0013] S3: Intelligent Recognition and Interactive Verification: The intelligent recognition module automatically loads the feature rule file of the crushed stone piles and performs pile recognition on the imported orthophoto. When the recognition accuracy is greater than 95%, it enters the interactive verification stage, where misidentified piles are manually removed, missing piles are added, and all piles are clustered and numbered in rows and columns to generate a crushed stone pile number map and attribute data table. If the recognition accuracy is less than 95%, it returns to the previous stage to supplement the number of labeled samples.

[0014] S4: Image Correction: The image correction module selects 3-6 ground control points within the orthophoto range, performs image correction on the gravel pile number map, and adds the X and Y coordinates of the pile body in the actual site coordinate system to the gravel pile attribute data table;

[0015] S5: Pile Location Detection: The pile location detection module imports an attribute data table containing the design number of the crushed stone pile, the design X coordinate, and the design Y coordinate, and performs PCA principal component analysis on the design pile location coordinates. It automatically calculates the deflection angle between the main direction of the pile group arrangement and the horizontal coordinate axis, and performs a rotation transformation on the design coordinates based on the deflection angle to align the pile group arrangement direction with the horizontal coordinate axis. Then, it performs cluster analysis on the rotated Y coordinates to automatically divide the row and column numbers of the design piles.

[0016] Using row and column number matching and pile center distance as the dual criteria, the actual piles and design piles are matched, the pile position deviation is calculated simultaneously, and the missing piles are identified.

[0017] In some embodiments, the process of pairing measured piles with designed piles based on row and column number matching and pile center distance verification, simultaneously calculating pile position deviations and identifying missing piles, includes:

[0018] Match the design piles and measured piles with the same row and column numbers, and then calculate the straight-line distance between their centers. When the center distance is less than 0.3 times the design pile diameter, it is judged as a qualified pile; when the center distance is greater than 0.3 times the design pile diameter but less than 1.0 times the design pile diameter, it is judged as a pile with excessive deviation; when the center distance is greater than 1.0 times the design pile diameter but less than 1.2 times the design pile diameter, it is judged as a pile with excessive deviation.

[0019] For a design pile that fails to match, if there is a measured pile within a range of 1.2 times the design pile diameter around it, it is determined to be a pile with excessive deviation; otherwise, it is determined to be a missing pile.

[0020] In some embodiments, the functional module further includes an output module, which automatically outputs a pile position deviation statistical report, a pile position deviation result plan view, and a data report.

[0021] In the specific implementation process, the pile position deviation statistical report includes indicators such as total number of piles, number of successfully matched piles, number of missing piles, average deviation, maximum deviation, minimum deviation, standard deviation of deviation, and pass rate.

[0022] The pile position deviation result plan uses green circles to indicate qualified piles, orange circles to indicate piles with excessive deviation, purple circles to indicate piles with excessive deviation, and red circles to indicate missing piles. The total number of piles, the number of missing piles, and the pile position deviation qualification rate are also marked on the drawing. In specific implementation, different colors can be used to mark qualified piles, piles with excessive deviation, piles with excessive deviation, and missing piles according to engineering practices. Those skilled in the art will understand this, and it will not be elaborated further here.

[0023] In some embodiments, the data report records the measured number, measured X / Y coordinates, matching design number, design X / Y coordinates, row and column numbers, pile position deviation, and qualification judgment result line by line.

[0024] In some embodiments, when labeling samples in step S2, the selected samples uniformly cover the piles in the orthophoto image with different lighting, grayscale, image edges and image center areas.

[0025] In some implementation sets, step S2 extracts six types of feature parameters for each sample, including blue channel value B, green channel value G, red channel value R, grayscale value, contour area, and roundness, specifically including:

[0026] Calculate the average values ​​of the red, green, and blue components of the local image to obtain the blue channel value B, green channel value G, and red channel value R;

[0027] After converting the color image to grayscale, the average brightness is calculated to obtain the grayscale value.

[0028] Call cv2.contourArea to calculate the pixel area of ​​the gravel pile region to obtain the outline area;

[0029] Get the perimeter of the outline using cv2.arcLength, and calculate it using the formula: Circularity = 4π × Area ÷ Perimeter 2 To achieve roundness;

[0030] Based on the feature values ​​of the labeled samples, a valid range of values ​​is automatically generated, including:

[0031] The range of blue channel value B, green channel value G, red channel value R, and grayscale value is: minimum value - maximum value of the sample;

[0032] Contour area range: minimum value × 0.7 to maximum value × 1.4;

[0033] Circularity range: minimum value × 0.85 to maximum value × 1.0.

[0034] In some embodiments, the intelligent recognition module in step S3 automatically loads the feature rule file of the crushed stone pile and performs pile body recognition on the imported orthophoto image, including:

[0035] (1) Preprocessing the orthophoto image to be identified, including:

[0036] Grayscale conversion: Call cv2.cvtColor to convert to grayscale;

[0037] Gaussian filtering: Call cv2.GaussianBlur to remove noise;

[0038] Binarization: Use cv2.threshold to convert the image to black and white using the Otsu threshold;

[0039] (2) Call cv2.findContours to find all contours;

[0040] (3) Calculate the area and roundness of each contour for preliminary screening, and directly remove contours that are not within the set range;

[0041] (4) For the remaining candidate regions, extract six-dimensional features again: R, G, B, gray level, contour area, and circularity.

[0042] (5) All features must meet the set range to be considered as crushed stone piles; any feature that exceeds the range will be considered as interference and removed.

[0043] (6) Spatial deduplication: Traverse all identification points and calculate the squared Euclidean distance d between two points. 2 = (x1−x2) 2 + (y1−y2)2 Where, if d 2 < (adaptive radius × 2) 2 They are considered as the same stake, and only one point is retained.

[0044] In some embodiments, the image correction module in step S4 selects 3-6 ground control points within the orthophoto range, performs image correction on the gravel pile numbering map, and adds the X and Y coordinates of the pile body in the actual site coordinate system to the gravel pile attribute data table, including:

[0045] (1) In the image correction window, the user clicks the control point pixel position in sequence and enters the corresponding site real coordinates (X, Y).

[0046] (2) The transformation matrix is ​​calculated using normalized least squares and affine transformation:

[0047] (2.1) Normalize the pixel coordinates and the true coordinates respectively (subtract the mean and divide by the standard deviation);

[0048] (2.2) Solve for the affine transformation matrix (6 degrees of freedom);

[0049] (2.3) If there are 3 control points, use exact affine transformation; if there are more than 3 control points, use least squares fitting.

[0050] (3) Apply transformation to the pixel coordinates of all identified crushed stone piles to obtain the X and Y coordinates in the actual coordinate system of the site.

[0051] This invention also provides an intelligent detection system for the position deviation of crushed stone piles, comprising:

[0052] The input module is used to import the orthophotos of the acquired crushed stone pile construction area;

[0053] The sample annotation module is used to calculate the average diameter and row and column spacing of the gravel piles in the orthophoto. Using the circular frame corresponding to the calculated average diameter as the annotation carrier, the gravel piles in the orthophoto are randomly and manually annotated. The number of annotated samples is no less than 60. Six types of feature parameters are extracted for each sample: including blue channel value B, green channel value G, red channel value R, gray value, contour area, and circularity. The module also adaptively calculates the binarization threshold based on the Otsu algorithm to generate a gravel pile feature rule file.

[0054] The intelligent recognition module is used to load the feature rule file of the crushed stone piles and identify the piles in the imported orthophoto. When the recognition accuracy is greater than 95%, it enters the interactive verification stage, where misidentified piles are manually removed, missing piles are added, and all piles are clustered and numbered in rows and columns to generate a crushed stone pile number map and attribute data table. If the recognition accuracy is less than 95%, it returns to the previous stage to supplement the number of labeled samples.

[0055] The image correction module is used to select 3-6 ground control points within the orthophoto range, perform image correction on the numbered map of the crushed stone piles, and add the X and Y coordinates of the piles in the actual coordinate system of the site to the crushed stone pile attribute data table;

[0056] The pile location detection module is used to import attribute data tables containing the design number of crushed stone piles, design X coordinate, and design Y coordinate, and perform PCA principal component analysis on the design pile location coordinates. It automatically calculates the deflection angle between the main direction of the pile group arrangement and the horizontal coordinate axis, and performs a rotation transformation on the design coordinates based on the deflection angle to align the pile group arrangement direction with the horizontal coordinate axis. Then, it performs cluster analysis on the rotated Y coordinates to automatically divide the row and column numbers of the design piles.

[0057] The output module is used to automatically output statistical reports on pile position deviations, planar diagrams of pile position deviation results, and data reports.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] This invention is based on a six-dimensional feature joint decision mechanism, which simultaneously extracts six-dimensional features: blue (B), green (G), red (R), grayscale, contour area, and circularity. Through multi-dimensional joint threshold decision, it significantly reduces the false detection rate and false negative rate in construction site environments such as complex lighting, surface debris, and pile obstruction.

[0060] This invention uses interactive measurement to obtain the average radius and row spacing of the crushed stone piles, and automatically calculates the binarization threshold based on the Otsu algorithm. Compared with a globally fixed threshold, it can dynamically adapt to images under different lighting conditions, improve the robustness of foreground segmentation, and eliminate the need for repeated manual parameter adjustments. The process can be directly reused in different sites, solving the problems of traditional methods relying on human experience and having poor generalization ability.

[0061] In view of the actual situation that the crushed stone pile group may rotate arbitrarily, the present invention performs principal component analysis (PCA) on the design coordinate set, automatically calculates the deflection angle between the main direction of the overall pile group and the horizontal coordinate axis, and performs rotation transformation on the coordinates accordingly. No manual input of angle or number of rows and columns is required, and it is applicable to rectangular or square pile layout sites with any orientation.

[0062] This invention first automatically assigns the row and column numbers of the designed piles and the measured piles through clustering, and then performs nearest neighbor distance matching within the same row and column number. This effectively solves the problem of one-to-one matching of pile positions in large-scale regular pile layouts (such as full-span piles) and avoids mismatches caused by blind global search.

[0063] This invention not only distinguishes between qualified and unqualified, but also further subdivides unqualified into "deviation exceeding the limit", "deviation too large" and "missing pile", comprehensively covering all kinds of quality problems that may occur in construction and acceptance, and the output results directly meet the acceptance requirements of the "Standard for Acceptance of Construction Quality of Building Foundation Engineering" (GB 50202-2018). Attached Figure Description

[0064] Figure 1 This is a flowchart of an embodiment of the intelligent detection method for pile position deviation of crushed stone piles according to the present invention;

[0065] Figure 2 This is a structural block diagram of the intelligent detection system for crushed stone pile position deviation of the present invention;

[0066] Figure 3 This is an orthophoto of a stone pile captured by a drone in Embodiment 1 of the present invention. The white dots in the image represent the stone piles.

[0067] Figure 4 A schematic diagram of the visual operation interface of the functional module constructed in Embodiment 1 of the present invention;

[0068] Figure 5 This is a schematic diagram of measuring the radius of a crushed stone pile during manual marking, as shown in Embodiment 1 of the present invention.

[0069] Figure 6 This is a schematic diagram of measuring the spacing between rows of crushed stone piles during manual marking, according to Embodiment 1 of the present invention.

[0070] Figure 7 This is a schematic diagram of the structure of manually labeled samples in Embodiment 1 of the present invention. In this diagram, 85 samples are manually labeled.

[0071] Figure 8 This is a schematic diagram illustrating how the intelligent identification module initially identified 594 crushed stone piles in Embodiment 1 of the present invention.

[0072] Figure 9 This is a diagram of the crushed stone pile numbers obtained by the intelligent identification module in Embodiment 1 of the present invention;

[0073] Figure 10 This is a schematic diagram of the image correction module correction process according to Embodiment 1 of the present invention;

[0074] Figure 11 This is a table showing the correction results of the image correction module in Embodiment 1 of the present invention;

[0075] Figure 12 This is a schematic diagram of the pile location detection module during detection according to Embodiment 1 of the present invention;

[0076] Figure 13 This is the pile position deviation statistical report output by Embodiment 1 of the present invention;

[0077] Figure 14 This is a plan view of the pile position deviation results output in Embodiment 1 of the present invention;

[0078] Figure 15 This is a data report output by Embodiment 1 of the present invention. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are some, but not all, of the embodiments of the present invention. The following embodiments are only for specific illustration of the implementation methods of the present invention and do not limit the scope of protection of the present invention.

[0080] In embodiments of the invention, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0081] In existing technologies, image recognition, coordinate transformation, and deviation calculation are typically applied as independent functional modules in different scenarios. This invention, for the first time, deeply integrates six-dimensional joint decision-making, PCA autonomous orientation, dual matching, four-level judgment, and full-process automation to produce synergistic technical effects. The technical solution of this invention is not a simple patchwork of independent modules, but rather forms an organic whole that solves a specific problem. Specifically, "interactive measurement to obtain prior parameters" provides necessary and adaptive high-precision initial values ​​for subsequent "cluster analysis to divide rows and columns," eliminating the unreliability of manually set parameters; the synergy between "Otsu adaptive threshold" and "six-dimensional feature joint decision-making" ensures the integrity and accuracy of extracting pile contours in complex backgrounds; and the synergy between "PCA autonomous orientation" and "dual matching strategy" solves the problem of coordinate alignment under arbitrary pile group rotation and the unique pairing challenge under large-scale dense pile layout. Finally, the refined output of the "four-level quality judgment system" directly addresses the rigid requirements of engineering acceptance. This series of interdependent technologies and hierarchical data transmissions together form an unprecedented fully automated detection chain. Its overall detection efficiency and robustness far exceed the simple extension of any single technology, and it has significant creative and practical application value.

[0082] This invention addresses the problem of inaccurate pile identification in complex construction site environments by constructing a complete technology chain: six-dimensional features → adaptive thresholding → PCA orientation → dual matching → four-level judgment. The absence of any single feature significantly degrades overall performance, forming an inseparable organic whole. The resulting effect far exceeds the sum of the effects of individual technologies. For example, single image recognition cannot solve large-scale matching problems; single PCA cannot improve pile identification accuracy; and single matching algorithms cannot adapt to arbitrary rotation and complex backgrounds. This invention, through the aforementioned complete technology chain, solves the problem of inaccurate pile identification in complex construction site backgrounds, achieving automatic alignment of arbitrarily rotated pile groups and one-to-one matching of large-scale dense pile locations. The detection results can directly serve project acceptance, possessing extremely high industrial application value.

[0083] Combined with appendix Figure 1 The intelligent detection method for pile position deviation of crushed stone piles of the present invention includes:

[0084] S1: Image Acquisition and Construction Functional Module for Crushed Stone Piles: Acquire orthophotos of the crushed stone pile construction area; construct functional modules based on the OpenCV image processing library and Python language, including a sample annotation module, an intelligent recognition module, an image correction module, and a pile location detection module.

[0085] The orthophoto of the crushed stone pile construction area was obtained by taking pictures with a drone. As a preferred method of the present invention, the ground resolution of the drone aerial image is 2 cm / pixel, that is, GSD<2 cm.

[0086] S2: Sample Labeling and Feature Rule File Generation: The average diameter and row-column spacing of the gravel piles in the orthophoto are calculated using the sample labeling module. Using the circular frame corresponding to the calculated average diameter as the labeling carrier, the gravel piles in the orthophoto are randomly and manually labeled. The number of labeled samples is no less than 60. Six types of feature parameters are extracted for each sample: including blue channel value B, green channel value G, red channel value R, gray value, contour area, and circularity. The binarization threshold is adaptively calculated based on the Otsu algorithm to generate the gravel pile feature rule file.

[0087] In the process of labeling samples, the selected samples uniformly cover the piles in the orthophoto with different lighting, gray levels, image edges and image center areas.

[0088] In the specific implementation process, the sample annotation module imports one or more sample images. Annotations are made by left-clicking on the center of the gravel pile, and a circular drawing function is called to generate a red circular marker at the pile location. Right-clicking undoes the previous annotation, facilitating corrections. Pressing the spacebar completes the annotation of the current image, with a minimum of 60 annotated samples. This 60-sample count is an empirical value, verified in multiple sites to cover over 90% of the six-dimensional feature value domain. When the lighting conditions at a new site vary significantly, it is recommended to increase the number of samples to 100-120. Using the annotated points as the center, a local image is cropped (cropping radius = adaptive radius × 2), and six core features are automatically calculated.

[0089] Among them, the adaptive radius is obtained through interactive measurement: the user drags a box to select the edge of the crushed stone pile, and the pixel radius is automatically calculated (at least 10 samples are collected and the average value is taken); the row spacing is obtained through interactive measurement: the user clicks the center of the corresponding crushed stone piles in two adjacent rows in turn, and the vertical pixel distance is automatically calculated.

[0090] Six-dimensional features are used to distinguish gravel piles from the background, rocks, and shadows. By expanding the feature range by a certain proportion, the adaptability and robustness of the model under different lighting and terrain conditions are improved. The joint decision rule for the six-dimensional features: This invention adopts a "AND logic" joint decision, meaning that a candidate target must simultaneously satisfy the feature value range of all six-dimensional features to be determined as a gravel pile; if any feature exceeds the range, it is determined as an interfering target and eliminated. Parallel threshold decision is used among the six-dimensional features without weights to ensure that the decision rule is transparent and reproducible.

[0091] S3: Intelligent Recognition and Interactive Verification: The intelligent recognition module automatically loads the feature rule file of the crushed stone piles and performs pile recognition on the imported orthophoto image to be recognized. When the recognition accuracy is greater than 95%, it enters the interactive verification stage, where misidentified piles are manually removed, missing piles are added, and all piles are clustered and numbered in rows and columns to generate a crushed stone pile number map and attribute data table. If the recognition accuracy is less than 95%, it returns to the previous stage to supplement the number of labeled samples.

[0092] In the specific implementation process, the intelligent recognition module automatically loads the feature rule file of the crushed stone piles and performs pile body recognition on the imported orthophoto images, including:

[0093] (1) Preprocessing the orthophoto image to be identified, including:

[0094] Grayscale conversion: Call cv2.cvtColor to convert to grayscale;

[0095] Gaussian filtering: Call cv2.GaussianBlur to remove noise;

[0096] Binarization: Use cv2.threshold to convert the image to black and white using the Otsu threshold;

[0097] (2) Call cv2.findContours to find all contours;

[0098] (3) Calculate the area and roundness of each contour for preliminary screening, and directly remove contours that are not within the set range;

[0099] (4) For the remaining candidate regions, extract six-dimensional features again: R, G, B, gray level, contour area, and circularity.

[0100] (5) All features must meet the set range to be considered as crushed stone piles; any feature that exceeds the range will be considered as interference and removed.

[0101] (6) Spatial deduplication: Traverse all identification points and calculate the squared Euclidean distance d between two points. 2 = (x1−x2) 2 + (y1−y2) 2 If d 2 < (adaptive radius × 2) 2 They are considered as the same stake, and only one point is retained.

[0102] The six-dimensional joint judgment significantly reduces the false detection rate and is suitable for complex construction site environments. Spatial deduplication avoids duplicate counting and ensures accurate pile quantity statistics.

[0103] The accuracy assessment specifically includes:

[0104] Randomly select no less than 100 recognition points (or 20% of all recognition points, whichever is larger) from the recognition results of the imported orthophotos, and automatically compare them with the labeled samples (or manually verify them quickly) to count the number of correctly recognized, misrecognized, and missed recognitions.

[0105] The accuracy rate is estimated using the following formula: Accuracy rate = (Number of correctly identified items ÷ (Number of correctly identified items + Number of incorrectly identified items + Number of missed identification items)) × 100%;

[0106] If the estimation accuracy is ≥ 95%, proceed to interactive verification;

[0107] If the estimated accuracy is less than 95%, return to step S2, increase the number of labeled samples, and retrain until the accuracy requirement is met.

[0108] In practice, interactive verification includes the following:

[0109] (1) A pop-up manual review window will appear, displaying the intelligent recognition results;

[0110] (2) Users can perform the following operations: Left click: Add a marker point at the missed pile location; Right click: Delete the incorrectly identified pile location (the program automatically finds the nearest marker point and deletes it);

[0111] (3) Press the space bar to complete the review;

[0112] (4) Perform Y-coordinate clustering based on the row spacing baseline value, and reorder all the verified points in the order from top to bottom and from left to right;

[0113] (5) Assign unique numbers 1, 2, 3... sequentially to generate a sequence numbering diagram and table of crushed stone piles, including the sequence number, image X coordinate, image Y coordinate, row number, and column number attributes.

[0114] Interactive verification ensures that the accuracy of identification meets the project acceptance requirements. The numbering rules fully conform to the inventory habits of engineering personnel, facilitating on-site verification.

[0115] S4: Image Correction: The image correction module selects 3-6 ground control points within the orthophoto range, performs image correction on the numbered map of the crushed stone piles, and adds the X and Y coordinates of the pile body in the actual coordinate system of the site to the crushed stone pile attribute data table.

[0116] In the specific implementation process, ground control points should be evenly distributed in the four corners and central area of ​​the image, covering at least 80% of the image area; it is prohibited to select areas with obvious image edge distortion (usually within 5% of the width of the surrounding area); control points should be selected from obvious surface landmarks (such as road markings, permanent features).

[0117] After correction, calculate the root mean square error (RMSE) of each control point. An RMSE ≤ 1 pixel (for this scheme, a 2cm resolution corresponds to an actual error ≤ 2cm) is considered acceptable. If the RMSE exceeds the limit, the number of control points should be increased or the points should be reselected until the accuracy requirements are met.

[0118] The image correction module selects 3-6 ground control points within the orthophoto range to perform image correction on the gravel pile numbering map, and adds the X and Y coordinates of the pile body in the actual site coordinate system to the gravel pile attribute data table, including:

[0119] (1) In the image correction window, the user clicks the control point pixel position in sequence and enters the corresponding site real coordinates (X, Y).

[0120] (2) The transformation matrix is ​​calculated using normalized least squares and affine transformation:

[0121] (2.1) Normalize the pixel coordinates and the true coordinates respectively (subtract the mean and divide by the standard deviation);

[0122] (2.2) Solve for the affine transformation matrix (6 degrees of freedom);

[0123] (2.3) If there are 3 control points, use exact affine transformation; if there are more than 3 control points, use least squares fitting.

[0124] (3) Apply transformation to the pixel coordinates of all identified gravel piles to obtain the X and Y coordinates in the true coordinate system of the site;

[0125] (4) Calculate and display the transformation accuracy (root mean square error RMSE, maximum error, minimum error).

[0126] The ground control point configuration can be saved as a JSON file and loaded and reused the next time it runs.

[0127] Affine transformations can correct rotation, translation, scaling, and slight tilt distortions in UAV imagery. After image correction, the gravel pile data table is updated with X and Y coordinate attributes in the actual site coordinate system.

[0128] S5: Pile Location Detection: The pile location detection module imports an attribute data table containing the design number of the crushed stone pile, the design X coordinate, and the design Y coordinate, and performs PCA principal component analysis on the design pile location coordinates. It automatically calculates the deflection angle between the main direction of the pile group arrangement and the horizontal coordinate axis, and performs a rotation transformation on the design coordinates based on the deflection angle to align the pile group arrangement direction with the horizontal coordinate axis. Then, it performs cluster analysis on the rotated Y coordinates to automatically divide the row and column numbers of the design piles.

[0129] Using row and column number matching and pile center distance as the dual criteria, the actual piles and design piles are matched, the pile position deviation is calculated simultaneously, and the missing piles are identified.

[0130] In the specific implementation process, before performing PCA principal component analysis, outliers in the design coordinate point cloud are first removed. The average distance and standard deviation of all design coordinate points to their nearest neighbors are calculated, and outliers (such as isolated edge points or incorrectly labeled points) whose average distance to neighboring points exceeds three times the standard deviation are removed. The removed outliers are marked separately, and row and column numbers are subsequently assigned manually.

[0131] PCA principal component analysis was performed on the design coordinate point cloud after outlier removal to obtain the overall layout direction and deflection angle. All design coordinates were then rotated to a horizontal coordinate system (X-axis along the layout direction, Y-axis perpendicular to the layout direction).

[0132] The next step involves performing cluster analysis on the rotated Y-coordinate to automatically assign row and column numbers to the designed piles, specifically including:

[0133] After rotating the Y coordinates, the actual line spacing is calculated using the interquartile range method, and Y coordinates are clustered and divided into lines with a threshold of 0.5 times the line spacing (i.e., if the difference between the Y coordinates of adjacent points exceeds this threshold, a new line is created).

[0134] Within each row, column numbers are assigned based on the X-coordinate, and a design table containing the design sequence number, number, coordinate, and row and column numbers is automatically output.

[0135] The PCA rotation parameters of the design coordinates are reused, and the measured coordinates are rotated to the same coordinate system and then assigned row and column numbers.

[0136] In some embodiments, the process of pairing measured piles with designed piles based on row and column number matching and pile center distance verification, simultaneously calculating pile position deviations and identifying missing piles, includes:

[0137] Match the design piles and measured piles with the same row and column numbers, and then calculate the straight-line distance between their centers. When the center distance is less than 0.3 times the design pile diameter, it is judged as a qualified pile; when the center distance is greater than 0.3 times the design pile diameter but less than 1.0 times the design pile diameter, it is judged as a pile with excessive deviation; when the center distance is greater than 1.0 times the design pile diameter but less than 1.2 times the design pile diameter, it is judged as a pile with excessive deviation.

[0138] For a design pile that fails to match, if there is a measured pile within a range of 1.2 times the design pile diameter around it, it is determined to be a pile with excessive deviation; otherwise, it is determined to be a missing pile.

[0139] The system prioritizes matching pile pairs (measured piles and design piles) with identical row and column numbers and the smallest spatial distance. Successfully matched pile pairs are categorized by distance as qualified piles, piles with excessive deviation, and piles with excessive deviation. For unmatched pile positions, a search is conducted for measured piles within a range of 1.2 times the design pile diameter. If a measured pile is found, it is marked as a pile with excessive deviation; otherwise, it is considered a missed pile.

[0140] PCA principal direction analysis requires no manual input of row numbers, column numbers, or arrangement angles. It is based entirely on data-driven automatic calculation and is applicable to crushed stone pile areas with any arrangement direction.

[0141] In some embodiments, the functional module further includes an output module, which automatically outputs a pile position deviation statistical report, a pile position deviation result plan view, and a data report.

[0142] In the specific implementation process, the pile position deviation statistical report includes indicators such as total number of piles, number of successfully matched piles, number of missing piles, average deviation, maximum deviation, minimum deviation, standard deviation of deviation, and pass rate.

[0143] The pile position deviation result plan uses green circles to indicate qualified piles, orange circles to indicate piles with excessive deviation, purple circles to indicate piles with excessive deviation, and red circles to indicate missing piles. The total number of piles, the number of missing piles, and the pile position deviation qualification rate are also marked on the drawing. In specific implementation, different colors can be used to mark qualified piles, piles with excessive deviation, piles with excessive deviation, and missing piles according to engineering practices. Those skilled in the art will understand this, and it will not be elaborated further here.

[0144] In some embodiments, the data report records the measured number, measured X / Y coordinates, matching design number, design X / Y coordinates, row and column numbers, pile position deviation, and qualification judgment result line by line.

[0145] In the specific implementation process, the test parameters, matching statistics, deviation statistics, missing pile list and acceptance conclusion are displayed in the form of pop-up windows. It supports one-click opening of CSV files and result images for manual review and fine-tuning.

[0146] All output files are saved to the desktop by default. The test reports fully meet the project acceptance requirements and support manual review and fine-tuning, satisfying the final accuracy requirements for project acceptance.

[0147] In some implementation sets, step S2 extracts six types of feature parameters for each sample, including blue channel value B, green channel value G, red channel value R, grayscale value, contour area, and roundness, specifically including:

[0148] Calculate the average values ​​of the red, green, and blue components of the local image to obtain the blue channel value B, green channel value G, and red channel value R;

[0149] After converting the color image to grayscale, the average brightness is calculated to obtain the grayscale value.

[0150] Call cv2.contourArea to calculate the pixel area of ​​the gravel pile region to obtain the outline area;

[0151] Get the perimeter of the outline using cv2.arcLength, and calculate it using the formula: Circularity = 4π × Area ÷ Perimeter 2 To achieve roundness;

[0152] Based on the feature values ​​of the labeled samples, a valid range of values ​​is automatically generated (i.e., the range set by the intelligent recognition module when identifying piles from imported orthophotos), including:

[0153] The range of blue channel value B, green channel value G, red channel value R, and grayscale value is: minimum value - maximum value of the sample;

[0154] Contour area range: minimum value × 0.7 to maximum value × 1.4;

[0155] Circularity range: minimum value × 0.85 to maximum value × 1.0.

[0156] Six-dimensional feature parameters are used to distinguish gravel piles from the background, rocks, and shadows. By expanding the feature range by a certain proportion, the adaptability and robustness of the model under different lighting and terrain conditions are improved.

[0157] This invention is based on a six-dimensional feature joint determination rule: This scheme adopts "AND logic" joint decision-making, that is, a candidate target must simultaneously satisfy the feature value range of all six-dimensional features in order to be determined as a gravel pile; if any feature exceeds the range, it is determined as an interfering target and eliminated. Parallel threshold decision-making is used among the six-dimensional features without weighting, ensuring that the determination rule is transparent and reproducible.

[0158] Combined with appendix Figure 2 The present invention also provides an intelligent detection system for pile position deviation of crushed stone piles, comprising:

[0159] The input module is used to import the orthophotos of the acquired crushed stone pile construction area;

[0160] The sample annotation module is used to calculate the average diameter and row and column spacing of the gravel piles in the orthophoto. Using the circular frame corresponding to the calculated average diameter as the annotation carrier, the gravel piles in the orthophoto are randomly and manually annotated. The number of annotated samples is no less than 60. Six types of feature parameters are extracted for each sample: including blue channel value B, green channel value G, red channel value R, gray value, contour area, and circularity. The module also adaptively calculates the binarization threshold based on the Otsu algorithm to generate a gravel pile feature rule file.

[0161] The intelligent recognition module is used to load the feature rule file of the crushed stone piles and identify the piles in the imported orthophoto. When the recognition accuracy is greater than 95%, it enters the interactive verification stage, where misidentified piles are manually removed, missing piles are added, and all piles are clustered and numbered in rows and columns to generate a crushed stone pile number map and attribute data table. If the recognition accuracy is less than 95%, it returns to the previous stage to supplement the number of labeled samples.

[0162] The image correction module is used to select 3-6 ground control points within the orthophoto range, perform image correction on the numbered map of the crushed stone piles, and add the X and Y coordinates of the piles in the actual coordinate system of the site to the crushed stone pile attribute data table;

[0163] The pile location detection module is used to import attribute data tables containing the design number of crushed stone piles, design X coordinate, and design Y coordinate, and perform PCA principal component analysis on the design pile location coordinates. It automatically calculates the deflection angle between the main direction of the pile group arrangement and the horizontal coordinate axis, and performs a rotation transformation on the design coordinates based on the deflection angle to align the pile group arrangement direction with the horizontal coordinate axis. Then, it performs cluster analysis on the rotated Y coordinates to automatically divide the row and column numbers of the design piles.

[0164] The output module is used to automatically output statistical reports on pile position deviations, planar diagrams of pile position deviation results, and data reports.

[0165] Example 1

[0166] The project site is located in Mimu Town, Qingbaijiang District, Chengdu City, Sichuan Province. The site contains a layer of miscellaneous fill soil with a thickness of 4.0–14.9 m. The project employs crushed stone piles for composite foundation treatment. The crushed stone piles have a design diameter of 0.8 m and are arranged in a square, full-area layout with a pile spacing of 1.6 m. According to Table 4.9.4 of the "Standard for Acceptance of Construction Quality of Building Foundation Engineering" (GB 50202-2018), the limit for pile position deviation in composite foundations is 0.3 times the design pile diameter, corresponding to a limit of 0.24 m for this project. To verify the practical application effect of the detection method of this invention, a smart detection test of crushed stone pile position deviation was conducted in a typical area of ​​the site.

[0167] S1: Image Acquisition and Construction Functional Module for Crushed Stone Piles

[0168] After the crushed stone piles were constructed, DJI Flip drones were used to collect orthophotos of the site. The flight path was parallel to the main direction of the overall crushed stone pile layout. The collected images had a resolution better than 2cm, as shown in Figure 3. Simultaneously, based on the OpenCV image processing library, functional modules were developed using Python, including a sample annotation module, an intelligent recognition module, an image correction module, and a pile position detection module. A visual operation interface was also built, completing the intelligent detection platform for crushed stone pile position deviations, as shown in Figure 4.

[0169] S2: Sample annotation and feature rule file generation

[0170] The sample annotation module was activated, and the pixel radius of the gravel piles was interactively measured in the orthophoto: 10 gravel piles were randomly selected for sampling, and the average pixel radius of the gravel piles was measured to be 9 pixels. The measurement of the gravel pile radius is as follows: Figure 5 As shown; then, 5 groups of adjacent pile rows were randomly selected, and the average row spacing of the piles was measured to be 56.3 pixels. The row spacing of the crushed stone piles was measured as follows. Figure 6 As shown.

[0171] Circular bounding boxes with a radius of 9 pixels were used to manually label the gravel piles within the image. The labeling samples included piles with different brightness, grayscale, size, and shape, resulting in a total of 85 labeled samples. The manual labeling samples are shown below. Figure 7 As shown (marked with red circles). The system automatically extracts six types of feature parameters for each sample: B, G, R channel values, grayscale value, contour area, and circularity. The Otsu algorithm is used to adaptively calculate the image binarization threshold of 174, and finally generates a JSON format gravel pile feature rule file.

[0172] The feature values ​​for this sample are as follows: blue channel value 108-182, green channel value 120-191, red channel value 140-198, gray value 127-192, contour area 108-988, and roundness 0.12-1.0.

[0173] S3: Intelligent Recognition and Interactive Verification

[0174] The feature rule file generated in step S2 is loaded, and the orthophoto image to be tested is sequentially processed with grayscale conversion, Gaussian filtering, binarization, morphological closing operation, contour detection, multi-feature joint decision, and Euclidean distance deduplication to complete the preliminary identification of the gravel piles. This preliminary identification identified 594 gravel piles (the designed number of gravel piles is 595), and the identification accuracy was determined to be greater than 95%, proceeding to the interactive verification stage. The plan view of the gravel piles initially identified by the intelligent recognition module is shown below. Figure 8 As shown.

[0175] During manual verification, one misidentified target was removed and one missed pile was added, ultimately confirming 594 validly identified gravel piles, achieving an overall recognition rate of 99%. The system combined the pile's Y-coordinate with an average row spacing of 56.3 pixels to perform row clustering analysis, automatically assigning pile numbers from top to bottom and left to right, generating a gravel pile numbering map and attribute data table. The data table includes the pile serial number, original pixel X-coordinate, original pixel Y-coordinate, row number, and column number. The recognition area was divided into 17 rows and 35 columns. The gravel pile numbering map identified by the intelligent recognition module is shown below. Figure 9 As shown.

[0176] S4: Image Correction

[0177] The image correction module is activated. Within the orthophoto range, four ground control points with known local engineering coordinates are selected. The control points are then sequentially located in the image, and their corresponding real coordinates are entered. The image correction process is automatically completed, converting the center pixel coordinates of all gravel piles to the measured coordinates of the site engineering. After correction, two new columns, "Corrected X Coordinate" and "Corrected Y Coordinate," are added to the gravel pile attribute data table. The correction process and results are shown in Figure 10. Figure 11 As shown.

[0178] S5: Pile Location Detection

[0179] Import a CSV-formatted design coordinate file containing design numbers, design X-coordinates, and design Y-coordinates. The pile location detection module performs PCA principal component analysis on the design pile location coordinate set to calculate the deflection angle of the overall arrangement direction of the crushed stone pile group relative to the horizontal coordinate axis. Based on this angle, a rotation transformation is performed on the design coordinates to make the pile group arrangement direction parallel to the horizontal coordinate axis. Cluster analysis is then performed on the rotated Y-coordinates to automatically assign row and column numbers to the design piles.

[0180] The testing area was designed with a total of 595 crushed stone piles. The calculated pile group deflection angle was 64.88°. The piles were designed to be divided into 17 rows and 35 columns. The testing process was as follows: Figure 12 As shown.

[0181] The matching and status determination of piles are completed using a dual verification rule of row and column number matching and pile center distance: design piles and measured piles with the same row and column number are matched, and the straight-line distance between their centers is calculated; when the center distance is less than 0.3 times the design pile diameter, it is judged as a qualified pile; when the center distance is greater than 0.3 times the design pile diameter but less than 1.0 times the design pile diameter, it is judged as a pile with excessive deviation; for design piles that fail to match, if there are measured piles within 1.2 times the design pile diameter around them, they are marked as piles with excessive deviation (usually corresponding to a distance greater than 1.0 times); otherwise, they are judged as missing piles.

[0182] The designed pile diameter for this project is 0.8m, the threshold for pile position deviation exceeding the limit is 0.24m, the threshold for excessive deviation is 0.80m, and the threshold for determining missing piles is 0.96m.

[0183] S6: Output Report

[0184] After the test is completed, three types of results are automatically output: a statistical report on pile position deviation, a plan view of pile position deviation results, and a data report, as shown in Figures 13, 14, and 15, respectively.

[0185] The total number of piles in the tested area was 595, and 593 pairs of piles were successfully matched with the design piles. Among them, 279 piles were qualified, 314 piles had deviations exceeding the limit, 1 pile had excessive deviation, and 1 pile was missing. The overall pass rate for pile position deviation was 47%. The average pile position deviation for the entire area was 0.268m, the maximum deviation was 0.748m, the minimum deviation was 0.003m, and the standard deviation of the deviation was 0.138m.

[0186] The entire process of testing 595 crushed stone piles in this area, from image import to final report generation, took approximately 9 minutes. Using the traditional total station for manual pile-by-pile measurement would require two surveyors working continuously for about two days, and the traditional method often relies on sampling, failing to achieve full coverage of the entire area.

[0187] The following table compares this embodiment with existing manual RTK methods and traditional manual image recognition methods:

[0188]

[0189] The comparison table above clearly demonstrates that this invention can achieve 100% full inspection of crushed stone pile locations, with a pile identification rate of 99%. The location and specific deviation values ​​of piles exceeding deviation limits and missing piles can be visualized, making defect locations clear and intuitive. Compared to traditional manual inspection methods, this solution significantly improves acceptance efficiency, with fully digitized inspection data and visualized results, fully meeting the application requirements of intelligent quality supervision at construction sites.

[0190] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for intelligent detection of pile position deviation in crushed stone piles, characterized in that, include: S1: Orthophoto acquisition and construction of crushed stone piles: acquire orthophotos of the crushed stone pile construction area; construct functional modules based on the OpenCV image processing library and Python language, including sample annotation module, intelligent recognition module, image correction module and pile location detection module; S2: Sample Labeling and Feature Rule File Generation: The average diameter and row-column spacing of the gravel piles in the orthophoto are calculated using the sample labeling module. Using the circular frame corresponding to the calculated average diameter as the labeling carrier, the gravel piles in the orthophoto are randomly and manually labeled. The number of labeled samples is no less than 60. Six types of feature parameters are extracted for each sample: including blue channel value B, green channel value G, red channel value R, gray value, contour area, and circularity. The binarization threshold is adaptively calculated based on the Otsu algorithm to generate the gravel pile feature rule file. S3: Intelligent Recognition and Interactive Verification: The intelligent recognition module automatically loads the feature rule file of the crushed stone piles and performs pile recognition on the imported orthophoto. When the recognition accuracy is greater than 95%, it enters the interactive verification stage, where misidentified piles are manually removed, missing piles are added, and all piles are clustered and numbered in rows and columns to generate a crushed stone pile number map and attribute data table. If the recognition accuracy is less than 95%, it returns to the previous stage to supplement the number of labeled samples. S4: Image Correction: The image correction module selects 3-6 ground control points within the orthophoto range, performs image correction on the gravel pile number map, and adds the X and Y coordinates of the pile body in the actual site coordinate system to the gravel pile attribute data table; S5: Pile Location Detection: The pile location detection module imports an attribute data table containing the design number of the crushed stone pile, the design X coordinate, and the design Y coordinate, and performs PCA principal component analysis on the design pile location coordinates. It automatically calculates the deflection angle between the main direction of the pile group arrangement and the horizontal coordinate axis, and performs a rotation transformation on the design coordinates based on the deflection angle to align the pile group arrangement direction with the horizontal coordinate axis. Then, it performs cluster analysis on the rotated Y coordinates to automatically divide the row and column numbers of the design piles. Using row and column number matching and pile center distance as the dual verification criteria, the actual piles and design piles are matched, the pile position deviation is calculated simultaneously, and the missing piles are identified.

2. The intelligent detection method for pile position deviation of crushed stone piles according to claim 1, characterized in that, The process of pairing measured piles with designed piles based on row and column number matching and pile center distance verification, simultaneously calculating pile position deviations and identifying missing piles, includes: Match the design piles and measured piles with the same row and column numbers, and then calculate the straight-line distance between their centers. When the center distance is less than 0.3 times the design pile diameter, it is judged as a qualified pile; when the center distance is greater than 0.3 times the design pile diameter but less than 1.0 times the design pile diameter, it is judged as a pile with excessive deviation; when the center distance is greater than 1.0 times the design pile diameter but less than 1.2 times the design pile diameter, it is judged as a pile with excessive deviation. For a design pile that fails to match the design, if there is a measured pile within a range of 1.2 times the design pile diameter around it, it is determined to be a pile with excessive deviation; otherwise, it is determined to be a missing pile.

3. The intelligent detection method for pile position deviation of crushed stone piles according to claim 1, characterized in that, The functional module also includes an output module, which automatically outputs a pile position deviation statistical report, a pile position deviation result plan view, and a data report.

4. The intelligent detection method for pile position deviation of crushed stone piles according to claim 1, characterized in that, When labeling samples in step S2, the selected samples uniformly cover the piles in the orthophoto image with different lighting, grayscale, image edges and image center areas.

5. The intelligent detection method for pile position deviation of crushed stone piles according to claim 2, characterized in that, Step S2 extracts six feature parameters for each sample, including blue channel value B, green channel value G, red channel value R, grayscale value, contour area, and roundness, specifically including: (5.1) Calculate the average values ​​of the red, green and blue components of the local image respectively to obtain the blue channel value B, green channel value G and red channel value R; (5.2) After converting the color image to grayscale, calculate the average brightness to obtain the grayscale value; (5.3) Call cv2.contourArea to calculate the pixel area of ​​the gravel pile region to obtain the outline area; (5.4) Call cv2.arcLength to get the perimeter of the outline, and calculate it according to the formula: Circularity = 4π × Area ÷ Perimeter 2 To achieve roundness; (5.5) Based on the feature values ​​of the labeled samples, automatically generate a valid range of values, where: The range of blue channel value B, green channel value G, red channel value R, and grayscale value is: minimum value - maximum value of the sample; Contour area range: minimum value × 0.7 to maximum value × 1.4; Circularity range: minimum value × 0.85 to maximum value × 1.

0.

6. The intelligent detection method for pile position deviation of crushed stone piles according to claim 5, characterized in that, In step S3, the intelligent recognition module automatically loads the feature rule file of the crushed stone piles and performs pile body recognition on the imported orthophoto image, including: (1) Preprocessing the orthophoto image to be identified, including: Grayscale conversion: Call cv2.cvtColor to convert to grayscale; Gaussian filtering: Call cv2.GaussianBlur to remove noise; Binarization: Use cv2.threshold to convert the image to black and white using the Otsu threshold; (2) Call cv2.findContours to find all contours; (3) Calculate the area and roundness of each contour for preliminary screening, and directly remove contours that are not within the set range; (4) For the remaining candidate regions, extract six-dimensional features again: R, G, B, gray level, contour area, and circularity. (5) All features must meet the set range to be considered as crushed stone piles; any feature that exceeds the range will be considered as interference and removed. (6) Spatial deduplication: Traverse all identification points and calculate the squared Euclidean distance d between two points. 2 = (x1−x2) 2 + (y1−y2) 2 .

7. The intelligent detection method for pile position deviation of crushed stone piles according to claim 6, characterized in that, In step S4, the image correction module selects 3-6 ground control points within the orthophoto range to perform image correction on the gravel pile numbering map. It then adds the X and Y coordinates of the pile body in the actual site coordinate system to the gravel pile attribute data table, including: (1) In the image correction window, the user clicks the control point pixel position in sequence and enters the corresponding site real coordinates (X, Y). (2) The transformation matrix is ​​calculated using normalized least squares and affine transformation: (2.1) Normalize the pixel coordinates and the ground truth coordinates respectively; (2.2) Solve for the affine transformation matrix; (2.3) If there are 3 control points, use exact affine transformation; if there are more than 3 control points, use least squares fitting. (3) Apply transformation to the pixel coordinates of all identified crushed stone piles to obtain the X and Y coordinates in the actual coordinate system of the site.

8. An intelligent detection system for pile position deviation of crushed stone piles, comprising: The input module is used to import the orthophotos of the acquired crushed stone pile construction area; The sample annotation module is used to calculate the average diameter and row and column spacing of the gravel piles in the orthophoto. Using the circular frame corresponding to the calculated average diameter as the annotation carrier, the gravel piles in the orthophoto are randomly and manually annotated. The number of annotated samples is no less than 60. Six types of feature parameters are extracted for each sample: including blue channel value B, green channel value G, red channel value R, gray value, contour area, and circularity. The module also adaptively calculates the binarization threshold based on the Otsu algorithm to generate a gravel pile feature rule file. The intelligent recognition module is used to load the feature rule file of the crushed stone piles and identify the piles in the imported orthophoto. When the recognition accuracy is greater than 95%, it enters the interactive verification stage, where misidentified piles are manually removed, missing piles are added, and all piles are clustered and numbered in rows and columns to generate a crushed stone pile number map and attribute data table. If the recognition accuracy is less than 95%, it returns to the previous stage to supplement the number of labeled samples. The image correction module is used to select 3-6 ground control points within the orthophoto range, perform image correction on the numbered map of the crushed stone piles, and add the X and Y coordinates of the piles in the actual coordinate system of the site to the crushed stone pile attribute data table; The pile location detection module is used to import attribute data tables containing the design number of crushed stone piles, design X coordinate, and design Y coordinate, and perform PCA principal component analysis on the design pile location coordinates. It automatically calculates the deflection angle between the main direction of the pile group arrangement and the horizontal coordinate axis, and performs a rotation transformation on the design coordinates based on the deflection angle to align the pile group arrangement direction with the horizontal coordinate axis. Then, it performs cluster analysis on the rotated Y coordinates to automatically divide the row and column numbers of the design piles. The output module is used to automatically output statistical reports on pile position deviations, planar diagrams of pile position deviation results, and data reports.

Images