Intelligent identification and high-precision measurement method and system for deep foundation pit defects

Abstract

The present application relates to a kind of deep foundation pit defect intelligent identification and high-precision measurement method and system, its method includes using GNSS static observation station to calculate the point coordinates of photography total station instrument monitoring station;Photography total station instrument is calibrated;Photography total station instrument is continuously patrolled on photography total station instrument monitoring station, and the sample set of deep foundation pit concrete support is obtained, and based on small sample depth long tail self-learning algorithm, the defect detection model of depth foundation pit is trained on line;Online target image is obtained and the pixel coordinates of defect point are identified in real time using trained depth foundation pit defect detection model, and the offset angle of defect point and photography total station instrument horizontal axis is calculated;According to offset angle, photography total station instrument is aligned to the three-dimensional coordinate measurement of defect point, and the parameter measurement information of defect point is obtained.The present application guarantees the stability and high accuracy of measurement monitoring station, obtains high-precision positioning information and calculates defect scale, finally provides for construction party to repair in time.

Description

Technical Field

[0001] This invention relates to the field of municipal monitoring technology, and in particular to an intelligent identification and high-precision measurement method and system for deep foundation pit defects. Background Technology

[0002] In recent years, extensive construction engineering practice has demonstrated that deep foundation pit projects undergo varying degrees of deformation in their structure and surrounding environment throughout their entire construction lifecycle. The explosive growth of deep foundation pit construction projects in cities has increased the likelihood of safety accidents and the difficulty of operation and management. The thicker soft soil layers and geological features in riverside, coastal, and other riverine areas pose significant challenges to the safety of foundation pit projects in these regions. To minimize the impact of deep foundation pit projects on the stability of their structure and surrounding environment during construction, defect monitoring of deep foundation pits is receiving increasing attention from all parties involved in the project. Traditional manual foundation pit monitoring is time-consuming and labor-intensive, and cannot provide real-time monitoring of sudden deformations. In traditional foundation pit monitoring, monitoring personnel need to manually inspect and collect deformation data for each item. Monitoring methods mainly involve manually inspecting the monitored object, measuring its overall tilt, position, and dimensions using instruments such as crack gauges, inclinometers, handheld frequency meters, and steel rulers. Although many automated monitoring technologies have emerged in recent years, such as measurement robots and Global Navigation Satellite Systems (GNSS), no single monitoring model can cover the entire lifecycle of deep foundation pits and all monitoring items. Work can only be carried out by combining manual and automated instruments, which still has unresolved drawbacks: data accuracy, timeliness, and completeness are all compromised; errors accumulate and increase; real-time deformation data cannot be obtained under adverse weather conditions; a large amount of maintenance manpower and equipment is consumed; the level of intelligence and automation is insufficient; monitoring time is limited; and costs are high and efficiency is low. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide an intelligent identification and high-precision measurement method and system for deep foundation pit defects, which addresses the shortcomings of the prior art.

[0004] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A method for intelligent identification and high-precision measurement of defects in deep foundation pits, comprising the following steps:

[0005] Using a GNSS static observation station, the coordinates of the photogrammetric total station monitoring station are calculated using the corner resection method.

[0006] A total photoelectric station is used to acquire calibration image information of a pre-set calibration plate, and the total photoelectric station is calibrated using the calibration image information and the coordinates of the monitoring station points.

[0007] The total station is used to continuously cruise at the total station monitoring station to obtain a sample set of concrete support for deep foundation pits, and the defect detection model of deep foundation pits is trained online based on the small sample depth long tail self-learning algorithm.

[0008] The target image is acquired online and the pixel coordinates of the defect points are identified in real time using the trained deep pit defect detection model. The offset angle between the defect points and the horizontal axis of the total station is calculated based on the pixel coordinates of the defect points.

[0009] Based on the offset angle, the total station is controlled to continuously perform three-dimensional coordinate measurements on the defect point to obtain the parameter measurement information of the defect point.

[0010] The beneficial effects of this invention are as follows: The intelligent identification and high-precision measurement method for deep foundation pit defects of this invention analyzes the stability of the monitoring station by combining a GNSS static observation station with a total photogrammetric station, and corrects the coordinates of the monitoring station in a timely manner to ensure the stability and high precision of the measurement monitoring station. The total photogrammetric station automatically and periodically cruises to acquire defect images, extracts defects in real time using a deep long-tail self-learning model, and simultaneously calibrates the relationship between the camera and the three-dimensional coordinate system of the total photogrammetric station to obtain extrinsic parameters. Based on the extrinsic parameters, the angle and direction are calculated, and the total photogrammetric station is automatically controlled to align with the defect for three-dimensional coordinate measurement, obtain high-precision positioning information, and calculate the defect scale. Finally, this information is provided to the construction party for timely repair and elimination of safety hazards.

[0011] Based on the above technical solution, the present invention can be further improved as follows:

[0012] Further: The calculation of the point coordinates of the photogrammetric total station monitoring station using the corner resection method with a GNSS static observation station specifically includes the following steps:

[0013] Calculate the coordinates of the total station P based on the geometric positional relationship between the two preset GNSS static observation stations A and B and the total station reference monitoring station P, and calculate the standard error of the total station monitoring station P.

[0014] The coordinates of the total station P are corrected based on the aforementioned mean square error.

[0015] The beneficial effect of the above-mentioned further scheme is that by utilizing the geometric positional relationship between two GNSS static observation stations A and B and the photogrammetric total station reference monitoring station P, the position coordinates of the photogrammetric total station reference monitoring station P can be accurately calculated. Then, based on the position coordinates of the photogrammetric total station reference monitoring station P, the standard error can be calculated to correct the position coordinates of the photogrammetric total station monitoring station P, thereby ensuring the stability and high accuracy of the measurement monitoring station.

[0016] Further: Based on the preset geometric positional relationship between the two GNSS static observation stations A and B and the total station monitoring station P, the point coordinates of the total station reference monitoring station P are calculated, and the standard error of the total station monitoring station P is calculated. This specifically includes the following steps:

[0017] The side length S between the reference monitoring station P and the GNSS static observation station A, as well as the angle β between the reference monitoring station P and the two GNSS static observation stations A and B, were measured using a total photostation.

[0018] The distance S0 between two GNSS static observation stations A and B and the direction angle α between them are obtained by measurement. AB and the coordinates (X) of GNSS static observation station A A ,Y A The coordinates (Xp, Yp) of the reference monitoring station P of the total station are calculated using the following formula:

[0019] ∠B=arcsin(S sinβ / S0) (1)

[0020] ∠A=180―(∠B+β) (2)

[0021] α AP =α AB +∠A (3)

[0022] Xp = X A +S cosα AP (4)

[0023] Yp = Y A +S sinα AP (5)

[0024] Where ∠B is the angle between the line connecting GNSS static observation station B and the total station reference monitoring station P, and the line connecting the two GNSS static observation stations A and B; ∠A is the angle between the line connecting GNSS static observation station A and the total station reference monitoring station P, and the line connecting the two GNSS static observation stations A and B; α AP Let α be the azimuth angle between GNSS static observation station A and photogrammetric total station reference monitoring station P. AB The azimuth angle between two GNSS static observation stations A and B;

[0025] Substituting equation (3) into (4) and (5), and taking the differential of the side length S between the total station P and the GNSS static observation station A, and the angle β between the total station P and the two GNSS static observation stations A and B, we get:

[0026]

[0027]

[0028] Converting to the form of standard error, we get:

[0029]

[0030]

[0031] The mean square error m of the (Xp, Yp) of the reference monitoring station P of the photogrammetric total station is... p for:

[0032]

[0033] Simplifying and rearranging, we get:

[0034]

[0035] Where ρ represents the radian-second constant.

[0036] The beneficial effect of the above-mentioned further scheme is that the position coordinates of the total station P can be accurately calculated by using the geometric positional relationship between the two GNSS static observation stations A and B and the total station monitoring station P. Then, the standard error can be accurately solved based on the side length S between the total station P and the GNSS static observation station A and the included angle β between the total station P and the two GNSS static observation stations A and B, so as to achieve the correction of the position coordinates of the total station P.

[0037] Further: The step of using a total photogrammetric station to acquire calibration image information of a pre-set calibration plate, and then using the calibration image information and the coordinates of the monitoring station points of the total photogrammetric station to perform external parameter calibration of the total photogrammetric station specifically includes the following steps:

[0038] A total photostation is used to acquire calibration image information of a pre-set calibration board. The SIFT algorithm is used to extract the corner points of the checkerboard in the calibration image information, obtain the image coordinates of the checkerboard corner points, and match them with the corresponding object points.

[0039] The rotation vector between the image-side coordinates and the corresponding object-side coordinates is calculated using Zhang Zhengyou's calibration method.

[0040] The rotation vector is converted into a rotation matrix R using the Rodriguez transformation, thus completing the external parameter calibration of the total station.

[0041] The beneficial effects of the above-mentioned further scheme are: by extracting the image coordinates of the checkerboard corner points in the calibration image information using the SIFT algorithm, and matching them with the object points, the rotation vector between the image coordinates and the corresponding object coordinates can be accurately calculated through the coordinate difference between the two, and finally the calibration rotation matrix R can be obtained, thereby realizing the external parameter calibration of the total station.

[0042] Further: The step of using the total station to continuously cruise at the total station monitoring station to obtain a sample set of concrete support for deep foundation pits, and then training the deep foundation pit defect detection model online based on a small sample depth long-tail self-learning algorithm specifically includes the following steps:

[0043] A deep foundation pit defect detection model is established, and the model loss function L of the deep foundation pit defect detection model is calculated based on the labeled sample set obtained by the cruise. The parameters of the deep foundation pit defect detection model are updated based on the model loss function L.

[0044] The updated deep foundation pit defect detection model is used to detect foundation pit defects and obtain the defect bounding box of the foundation pit in the image.

[0045] The defect rectangle (box) is used as a prompt word and input into the SAM model for refined extraction of crack defects, thereby obtaining the edge line of the crack.

[0046] The beneficial effects of the above-mentioned further scheme are: by calculating the model loss function L of the deep foundation pit defect detection model through a pre-collected labeled sample set and updating the model, the foundation pit defects can be identified more accurately, and the defect rectangle box can be obtained. Finally, the crack defect is refined by using the SAM model to obtain the edge line of the crack, thus achieving accurate defect identification.

[0047] Further: The calculation of the model loss function L of the deep foundation pit defect detection model based on the labeled sample set obtained from the cruise specifically includes the following steps:

[0048] The supervised loss function L of the deep foundation pit defect detection model is calculated based on the labeled sample set obtained from the cruise. l The calculation formula is:

[0049] p l,k ,b l.k =PPYOLOE-k(x l (9)

[0050]

[0051]

[0052]

[0053] Where PPYOLOE-k(·) represents the k-th detection head, p l,k and b l,k Let Cls(·) and Reg(·) represent the target probability score and detection box output by the k-th detection head, respectively. Let Cls(·) and Reg(·) represent the defect classification and bounding box loss functions, respectively. and These represent the classification loss and bounding box regression loss for classifying foundation pit defects, respectively. l b l and y l These represent the input foundation pit image, foundation pit defect bounding box, and defect category, respectively.

[0054] For the unlabeled sample set D acquired during the cruise... U ={x u First, PPYOLOE-F was used to detect foundation pit defects in the unlabeled data:

[0055] p u,k ,b u,k =PPYOLOE-k(x u (13)

[0056] Where p u,k and b u,k Let b represent the target probability score and detection box output by the k-th detector, respectively. The detection boxes from multiple detectors are fused to obtain the pseudo-box b. u,f and its corresponding target probability score p u,f and category y u :

[0057]

[0058] p u,f =max(p u,mean (14)

[0059] y u =argmax(p u,mean (15)

[0060]

[0061] The statistical depth foundation pit defect detection model calculates the average probability score μ and standard deviation vector δ for each type of defect, and sets a pseudo-boundary filtering threshold for each category.

[0062]

[0063] in, μ represents the screening threshold of the k-th detector head PPYOLOE-k(·) for pit category c.c and δ c The table shows the mean μ and standard deviation vector δ of the probability scores for foundation pit category c, respectively;

[0064] Calculate the unsupervised loss function L u :

[0065]

[0066]

[0067]

[0068] in, This indicates that the k-th detection head is based on a threshold. For category y u The pseudo-boundaries are filtered out, and the target probability score p of the pseudo-boundary is... u,f Greater than or equal to the threshold The value is 1, and the pseudobox participates in the unsupervised loss function L. u The value is calculated and used in model training; otherwise, it is set to 0. Pseudoboxes do not participate in the calculation of the unsupervised loss function and do not participate in training.

[0069] Based on the supervision loss function L l and unsupervised loss function L u The model loss function L of the few-sample deep long-tail self-learning model is calculated using the following formula:

[0070] L = L l +L u (20)

[0071] That is, the model loss function L is the supervision loss function L l and unsupervised loss function L u sum.

[0072] The beneficial effect of the above further scheme is that it calculates the supervised loss function L of the deep foundation pit defect detection model using a pre-collected labeled sample set. l Then, the unsupervised loss function L is calculated using the given unlabeled sample set. u Furthermore, through the supervised loss function L l and unsupervised loss function L u The sum of these factors accurately determines the model loss function L, which is then used to update the model, allowing for more precise identification of foundation pit defects.

[0073] Further: The calculation of the offset angle between the defect point and the horizontal axis of the total station based on the pixel coordinates of the defect point specifically includes the following steps:

[0074] Calculate the two-dimensional offset angle (α, β) between the defect point and the principal axis of the total station camera based on the pixel coordinates of the defect point and the principal distance of the total station camera image:

[0075]

[0076] Where (x,y) are the pixel coordinates of the defect point, and f is the principal distance of the camera image, (x,y) c ,y c () represents the coordinates of the principal point;

[0077] Calculate the offset angle γ between the defect point and the horizontal axis of the total station based on the calibration parameters of the total station:

[0078]

[0079] Where R is the rotation matrix calibrated by the total photostation.

[0080] The beneficial effect of the above-mentioned further scheme is that the two-dimensional offset angle between the defect point and the principal axis of the camera of the total station can be accurately calculated by using the pixel coordinates of the defect point and the camera image principal distance of the total station. Then, based on the two-dimensional offset angle and the calibration parameters of the total station, the offset angle between the defect point and the horizontal axis of the total station can be calculated, which facilitates the subsequent control of the total station to continuously perform three-dimensional coordinate measurements on the defect point and accurately obtain the parameter measurement information of the defect point.

[0081] The present invention also provides an intelligent identification and high-precision measurement system for deep foundation pit defects, including a calculation module, a calibration module, a training module, an identification module and a measurement module;

[0082] The calculation module is used to calculate the point coordinates of the total station monitoring station using the corner resection method with GNSS static observation station.

[0083] The calibration module is used to acquire calibration image information of a pre-set calibration plate using a photogrammetric total station, and to perform external parameter calibration on the photogrammetric total station using the calibration image information.

[0084] The training module is used to continuously cruise the total station of the camera at the monitoring station to obtain a sample set of concrete support for deep foundation pits, and to train the deep foundation pit defect detection model online based on the small sample depth long tail self-learning algorithm.

[0085] The recognition module is used to acquire target images online and use the trained deep pit defect detection model to identify the pixel coordinates of defect points in real time, and calculate the offset angle between the defect point and the horizontal axis of the total station based on the pixel coordinates of the defect point.

[0086] The measurement module is used to control the total station to continuously perform three-dimensional coordinate measurements on the defect point according to the offset angle, and to obtain the parameter measurement information of the defect point.

[0087] The intelligent identification and high-precision measurement system for deep foundation pit defects of this invention corrects the coordinates of the station to ensure the stability and high precision of the measurement and monitoring station. It automatically acquires defect images through timed cruise of a total photostation, extracts defects in real time using a deep long-tail self-learning model, and simultaneously calibrates the relationship between the camera and the three-dimensional coordinate system of the total photostation to obtain extrinsic parameters. Based on the extrinsic parameters, the angle and direction are calculated, and the total photostation is automatically controlled to align with the defect for three-dimensional coordinate measurement, obtaining high-precision positioning information and calculating the defect scale. Finally, it provides the information to the construction party for timely repair and elimination of safety hazards.

[0088] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the intelligent identification and high-precision measurement method for deep foundation pit defects.

[0089] The present invention also provides an intelligent identification and high-precision measurement device for deep foundation pit defects, characterized in that: it includes a communication interface, a memory, a communication bus and a processor, wherein the processor, the communication interface and the memory communicate with each other through the communication bus;

[0090] The memory is used to store computer programs;

[0091] When the processor executes the program stored in the memory, it implements the steps of the intelligent identification and high-precision measurement method for deep foundation pit defects. Attached Figure Description

[0092] Figure 1 This is a flowchart illustrating an embodiment of the intelligent identification and high-precision measurement method for deep foundation pit defects according to the present invention.

[0093] Figure 2 This is a schematic diagram showing the geometric layout of two GNSS static observation stations A and B and the total station reference monitoring station P according to an embodiment of the present invention.

[0094] Figure 3 This is a schematic diagram of the structure of an intelligent identification and high-precision measurement system for deep foundation pit defects according to an embodiment of the present invention. Detailed Implementation

[0095] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0096] like Figure 1As shown, a method for intelligent identification and high-precision measurement of defects in deep foundation pits includes the following steps:

[0097] S1: Using a GNSS static observation station, the coordinates of the photogrammetric total station monitoring station are calculated using the corner resection method;

[0098] S2: Use a total photoelectric station to acquire calibration image information of a pre-set calibration plate, and use the calibration image information and the point coordinates of the total photoelectric station monitoring station to perform external parameter calibration on the total photoelectric station;

[0099] S3: Utilize the total station to continuously cruise at the total station monitoring station to obtain a sample set of concrete support for deep foundation pits, and train the deep foundation pit defect detection model for online defect identification based on the small sample depth long tail self-learning algorithm.

[0100] S4: Acquire the target image online and use the trained deep pit defect detection model to identify the pixel coordinates of the defect points in real time, and calculate the offset angle between the defect points and the horizontal axis of the total station based on the pixel coordinates of the defect points.

[0101] S5: Control the total station to continuously measure the three-dimensional coordinates of the defect point according to the offset angle, and obtain the parameter measurement information of the defect point.

[0102] The present invention provides an intelligent identification and high-precision measurement method for deep foundation pit defects. This method analyzes the stability of the monitoring station using a GNSS static observation station combined with a total photogrammetric station, and promptly corrects the station's coordinates to ensure stability and high precision. The total photogrammetric station automatically and periodically cruises to acquire defect images, extracts defects in real time using a deep long-tail self-learning model, and simultaneously calibrates the relationship between the camera and the total station's three-dimensional coordinate system to obtain extrinsic parameters. Based on these extrinsic parameters, the angle and direction are calculated, and the total photogrammetric station is automatically controlled to align with the defect and perform three-dimensional coordinate measurement. This obtains high-precision positioning information and calculates the defect's scale, ultimately providing the information to the construction team for timely repairs and to eliminate safety hazards.

[0103] In practice, due to continuous construction within or around the foundation pit, it is impossible to provide a completely stable bedrock point as a site for a total station monitoring device. Therefore, changes in the monitoring station's location are unavoidable during long-term monitoring. Instability at the monitoring station can damage the measurement benchmark, making long-term monitoring impossible. This invention utilizes a GNSS static observation station and employs the corner resection method to recalculate the coordinates of the measurement robot monitoring station for stability analysis and correction.

[0104] In one or more embodiments of the present invention, the calculation of the point coordinates of the photogrammetric total station monitoring station using the corner resection method with a GNSS static observation station specifically includes the following steps:

[0105] S11: Calculate the coordinates of the total station P based on the geometric positional relationship between the two preset GNSS static observation stations A and B and the total station reference monitoring station P, and calculate the standard error of the total station monitoring station P.

[0106] S12: Correct the position coordinates of the total station monitoring station P based on the aforementioned mean square error.

[0107] By utilizing the geometric positional relationship between two GNSS static observation stations A and B and the total station reference monitoring station P, the position coordinates of the total station reference monitoring station P can be accurately calculated. Then, based on the position coordinates of the total station reference monitoring station P, the standard error can be calculated to correct the position coordinates of the total station monitoring station P, thereby ensuring the stability and high accuracy of the measurement monitoring station.

[0108] In one or more embodiments of the present invention, the calculation of the point coordinates of the photogrammetric total station reference monitoring station P based on the geometric positional relationship between two preset GNSS static observation stations A and B and the photogrammetric total station monitoring station P, and the calculation of the standard error of the photogrammetric total station monitoring station P specifically include the following steps:

[0109] S111: Measure the side length S between the total station P and the GNSS static observation station A, as well as the angle β between the total station P and the two GNSS static observation stations A and B, using a total station photogrammetry.

[0110] like Figure 2 As shown, points A and B are both GNSS static observation stations. A coaxial 360° prism is installed on the observation station. The GNSS acquires the real-time coordinates of point A. The coaxial 360° prism is used for the measurement of the intersection of the corners of the total station. Point P is the monitoring station of the photogrammetric total station.

[0111] S112: Measure the distance S0 between two GNSS static observation stations A and B, and obtain the direction angle α between them. AB and the coordinates (X) of GNSS static observation station A A ,Y A The coordinates (Xp, Yp) of the reference monitoring station P of the total station are calculated using the following formula:

[0112] ∠B=arcsin(S sinβ / S0) (1)

[0113] ∠A=180―(∠B+β) (2)

[0114] α AP =α AB +∠A (3)

[0115] Xp = X A +S cosα AP (4)

[0116] Yp = Y A +S sinα AP (5)

[0117] Where ∠B is the angle between the line connecting GNSS static observation station B and the total station reference monitoring station P, and the line connecting the two GNSS static observation stations A and B; ∠A is the angle between the line connecting GNSS static observation station A and the total station reference monitoring station P, and the line connecting the two GNSS static observation stations A and B; α AP Let α be the azimuth angle between GNSS static observation station A and photogrammetric total station reference monitoring station P. AB The azimuth angle between two GNSS static observation stations A and B;

[0118] S113: Substituting equation (3) into (4) and (5), and taking the differential of the side length S between the total station reference monitoring station P and the GNSS static observation station A, and the angle β between the total station reference monitoring station P and the two GNSS static observation stations A and B, we get:

[0119]

[0120]

[0121] Converting to the form of standard error, we get:

[0122]

[0123]

[0124] The mean square error m of the (Xp, Yp) of the reference monitoring station P of the photogrammetric total station is... p for:

[0125]

[0126] Simplifying and rearranging, we get:

[0127]

[0128] Where ρ represents the radian-second constant.

[0129] By using the geometric positional relationship between two GNSS static observation stations A and B and the total station monitoring station P, the positional coordinates of the total station reference monitoring station P can be accurately calculated. Then, based on the side length S between the total station reference monitoring station P and the GNSS static observation station A, and the angle β between the total station reference monitoring station P and the two GNSS static observation stations A and B, the standard error can be accurately solved to achieve the correction of the positional coordinates of the total station reference monitoring station P.

[0130] After the total station is installed and leveled, the three-dimensional coordinate system can be in a highly accurate horizontal state. However, due to the effective manufacturing process, the image space coordinate system of the camera (CCD module) in the total station will be rotated and offset relative to the horizontal total station coordinate system. Therefore, it is necessary to perform external parameter calibration of the camera relative to the total station's three-dimensional coordinate system.

[0131] In one or more embodiments of the present invention, the step of using a total photoelectric station to acquire calibration image information of a pre-set calibration plate, and using the calibration image information and the coordinates of the monitoring station points of the total photoelectric station to perform external parameter calibration of the total photoelectric station specifically includes the following steps:

[0132] S21: Use a total station to acquire calibration image information of a pre-set calibration board, use the SIFT algorithm to extract the corner points of the checkerboard in the calibration image information, obtain the image coordinates of the checkerboard corner points, and match them with the corresponding object points.

[0133] Here, a high-precision calibration plate is first made. Flat, ripple-free glass is used as the substrate, and aluminum oxide is attached to the substrate as the surface material. The surface pattern is a black and white chessboard pattern with clear corner points. The chessboard pattern needs to have a certain width of blank space around it. Such a calibration plate ensures the coplanarity of the corner point height and has a good matte effect, so as to make the corner point detection accuracy in the image high. The object space coordinates of the corner points of the calibration plate are obtained by total station measurement, with an accuracy of millimeters.

[0134] To ensure the camera of the total station is aimed at the calibration board and acquires an image, the calibration board must roughly fill the entire field of view. The SIFT algorithm is used to extract the corner points of the checkerboard pattern in the image, obtain their image coordinates, and match the corresponding object coordinates.

[0135] S22: Calculate the rotation vector between the image-space coordinates and the corresponding object-space coordinates using Zhang Zhengyou's calibration method;

[0136] S23: Use the Rodriguez transformation to convert the rotation vector into a rotation matrix R, and complete the external parameter calibration of the total station.

[0137] The SIFT algorithm is used to extract the image coordinates of the checkerboard corner points in the calibration image information and match them with the object points. Then, the rotation vector between the image coordinates and the corresponding object coordinates is accurately calculated by the coordinate difference between the two, and finally the calibration rotation matrix R is obtained, thus realizing the external parameter calibration of the photogrammetric total station.

[0138] In one or more embodiments of the present invention, the step of using the total station to continuously cruise at the total station monitoring station to obtain a sample set of deep foundation pit concrete support, and training the deep foundation pit defect detection model online based on a small sample depth long-tail self-learning algorithm specifically includes the following steps:

[0139] S31: Establish a deep foundation pit defect detection model, calculate the model loss function L of the deep foundation pit defect detection model based on the labeled sample set obtained by the cruise, and update the parameters of the deep foundation pit defect detection model based on the model loss function L.

[0140] S32: Detect foundation pit defects using the updated depth foundation pit defect detection model and obtain the defect bounding box of the foundation pit in the image;

[0141] S33: Input the defect rectangle box as a prompt word into the SAM model to perform refined extraction of crack defects and obtain the edge line of the crack.

[0142] The model loss function L of the deep foundation pit defect detection model is calculated by pre-collecting a labeled sample set and the model is updated. This allows for more accurate identification of foundation pit defects, obtaining the defect rectangle box. Finally, the SAM model is used to refine the extraction of crack defects, obtain the crack edge line, and achieve accurate defect identification.

[0143] In one or more embodiments of the present invention, the step of calculating the model loss function L of the deep foundation pit defect detection model based on the labeled sample set obtained from the cruise specifically includes the following steps:

[0144] S311: Calculate the supervised loss function L of the deep foundation pit defect detection model based on the labeled sample set obtained from the cruise. l The calculation formula is:

[0145] p l,k ,b l.k =PPYOLOE-k(x l (9)

[0146]

[0147]

[0148]

[0149] Where PPYOLOE-k(·) represents the k-th detection head, p l,k and b l,k Let Cls(·) and Reg(·) represent the target probability score and detection box output by the k-th detection head, respectively. Let Cls(·) and Reg(·) represent the defect classification and bounding box loss functions, respectively. and The classification loss and bounding box regression loss are respectively used to classify foundation pit defects, x l b l and y l These represent the input foundation pit image, foundation pit defect bounding box, and defect category, respectively.

[0150] In practice, automatic control technology is used to drive a total photogrammetric station to continuously cruise at the total photogrammetric station monitoring station, covering a horizontal range of 360° and a vertical zenith distance of 45° to 135°, acquiring online images of the concrete support structure of deep foundation pits as a sample set. Using the collected detection dataset, which contains only a small amount of labeled data and is class-imbaled, a long-tail self-learning algorithm is employed to train a deep foundation pit defect detection model.

[0151] S312: For the unlabeled sample set D acquired during cruise... U ={x u First, PPYOLOE-F was used to detect foundation pit defects in the unlabeled data:

[0152] p u,k ,b u,k =PPYOLOE-k(x u (13)

[0153] Where p u,k and b u,k Let b represent the target probability score and detection box output by the k-th detector, respectively. The detection boxes obtained from multiple detectors are fused to obtain the pseudo-box b. u,f and its corresponding target probability score p u,f and category y u :

[0154]

[0155] p u,f =max(p u,mean (14)

[0156] y u =argmax(p u,mean (15)

[0157]

[0158] The PPYOLOE model is used for foundation pit defect detection. To address the issues of small number of foundation pit defect label samples and class imbalance (i.e., long-tail problem), three additional detection heads are added to the PPYOLOE model, and the detection model is trained using a few-sample deep long-tail learning algorithm.

[0159] S313: The statistical depth foundation pit defect detection model calculates the average probability score μ and standard deviation vector δ for each type of defect, and sets a pseudo-box filtering threshold for each category.

[0160]

[0161] in, μ represents the screening threshold of the k-th detector head PPYOLOE-k(·) for pit category c. c and δ c The table shows the mean μ and standard deviation vector δ of the probability scores for foundation pit category c, respectively;

[0162] Here, u = mean(p) l,* ), δ=std(p l,* ), where p l,* ={p l,k |k=1,2,3,4} represents the set of target probability scores output by the four detection heads of the deep foundation pit defect detection model on labeled data. Then, a filtering threshold for pseudo-boundaries is set for each category of each detection head.

[0163] S314: Calculate the unsupervised loss function L u :

[0164]

[0165]

[0166]

[0167] in, This indicates that the k-th detection head is based on a threshold. For category y u The pseudo-boundaries are filtered out, and the target probability score p of the pseudo-boundary is... u,f Greater than or equal to the threshold The value is 1, and the pseudobox participates in the unsupervised loss function L. u The value is calculated and used in model training; otherwise, it is set to 0. Pseudoboxes do not participate in the calculation of the unsupervised loss function and do not participate in training.

[0168] S315: Based on the supervised loss function L l and unsupervised loss function L uThe model loss function L of the few-sample deep long-tail self-learning model is calculated using the following formula:

[0169] L = L l +L u (20)

[0170] That is, the model loss function L is the supervision loss function L l and unsupervised loss function L u sum.

[0171] The supervised loss function L of the deep foundation pit defect detection model is calculated using a pre-collected set of labeled samples. l Then, the unsupervised loss function L is calculated using the given unlabeled sample set. u Furthermore, through the supervised loss function L l and unsupervised loss function L u The sum of these factors accurately determines the model loss function L, which is then used to update the model, allowing for more precise identification of foundation pit defects.

[0172] In one or more embodiments of the present invention, the step of calculating the offset angle between the defect point and the horizontal axis of the total station based on the pixel coordinates of the defect point specifically includes the following steps:

[0173] S41: Calculate the two-dimensional offset angle (α, β) between the defect point and the principal axis of the total station camera based on the pixel coordinates of the defect point and the principal distance of the total station camera image:

[0174]

[0175] Where (x,y) are the pixel coordinates of the defect point, and f is the principal distance of the camera image, (x,y) c ,y c () represents the coordinates of the principal point;

[0176] S42: Calculate the offset angle γ between the defect point and the horizontal axis of the total station based on the calibration parameters of the total station:

[0177]

[0178] Where R is the rotation matrix calibrated by the total photostation.

[0179] By using the pixel coordinates of the defect point and the camera image principal distance of the total station, the two-dimensional offset angle between the defect point and the principal optical axis of the total station is accurately calculated. Then, based on the two-dimensional offset angle and the calibration parameters of the total station, the offset angle between the defect point and the horizontal axis of the total station is calculated. This facilitates subsequent control of the total station to continuously perform three-dimensional coordinate measurements on the defect point, and accurately obtains the parameter measurement information of the defect point.

[0180] In this invention, after calculating the offset angle between the defect point and the horizontal axis of the total station, the total station is controlled to continuously perform three-dimensional coordinate measurements on the defect point according to the offset angle, and the positioning coordinates, length information, area information and other parameter measurement information of the defect point are obtained. The specific calculation is existing technology and will not be described in detail in this invention.

[0181] The intelligent identification and high-precision measurement method for deep foundation pit defects of the present invention can be used for settlement monitoring of extra-large irregular buildings that cannot be directly accessed, or buildings where it is inconvenient to install measuring equipment and markers, and has broad application prospects.

[0182] like Figure 3 As shown, the present invention also provides an intelligent identification and high-precision measurement system for deep foundation pit defects, including a calculation module, a calibration module, a training module, an identification module and a measurement module;

[0183] The calculation module is used to calculate the point coordinates of the total station monitoring station using the corner resection method with GNSS static observation station.

[0184] The calibration module is used to acquire calibration image information of a pre-set calibration plate using a photogrammetric total station, and to perform external parameter calibration on the photogrammetric total station using the calibration image information.

[0185] The training module is used to continuously cruise the total station of the camera at the monitoring station to obtain a sample set of concrete support for deep foundation pits, and to train the deep foundation pit defect detection model online based on the small sample depth long tail self-learning algorithm.

[0186] The recognition module is used to acquire target images online and use the trained deep pit defect detection model to identify the pixel coordinates of defect points in real time, and calculate the offset angle between the defect point and the horizontal axis of the total station based on the pixel coordinates of the defect point.

[0187] The measurement module is used to control the total station to continuously perform three-dimensional coordinate measurements on the defect point according to the offset angle, and to obtain the parameter measurement information of the defect point.

[0188] The intelligent identification and high-precision measurement system for deep foundation pit defects of this invention corrects the coordinates of the measurement and monitoring station to ensure its stability and high accuracy. It automatically acquires defect images through timed cruise of a total photostation, extracts defects in real time using a small-sample depth long-tail self-learning model, and simultaneously calibrates the relationship between the camera and the total station's three-dimensional coordinate system to obtain extrinsic parameters. Based on the extrinsic parameters, the angle and direction are calculated, and the total photostation is automatically controlled to align with the defect for three-dimensional coordinate measurement, obtaining high-precision positioning information and calculating the defect scale. Finally, this information is provided to the construction party for timely repair and elimination of safety hazards.

[0189] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the intelligent identification and high-precision measurement method for deep foundation pit defects.

[0190] The present invention also provides an intelligent identification and high-precision measurement device for deep foundation pit defects, characterized in that: it includes a communication interface, a memory, a communication bus and a processor, wherein the processor, the communication interface and the memory communicate with each other through the communication bus;

[0191] The memory is used to store computer programs;

[0192] When the processor executes the program stored in the memory, it implements the steps of the intelligent identification and high-precision measurement method for deep foundation pit defects.

[0193] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent identification and high-precision measurement of defects in deep foundation pits, characterized in that: Includes the following steps: use GNSS For static observation stations, the coordinates of the total station monitoring points are calculated using the corner resection method. A total photoelectric station is used to acquire calibration image information of a pre-set calibration plate, and the total photoelectric station is calibrated using the calibration image information and the coordinates of the monitoring station points. The total station is used to continuously cruise at the total station monitoring station to obtain a sample set of concrete support for deep foundation pits, and the defect detection model of deep foundation pits is trained online based on the small sample depth long tail self-learning algorithm. The target image is acquired online and the pixel coordinates of the defect points are identified in real time using the trained deep pit defect detection model. The offset angle between the defect points and the horizontal axis of the total station is calculated based on the pixel coordinates of the defect points. Based on the offset angle, the total station is controlled to continuously perform three-dimensional coordinate measurements on the defect point to obtain the parameter measurement information of the defect point. The process of using the total station to continuously cruise at the monitoring station to acquire a sample set of concrete support for deep foundation pits, and then training the online defect identification model for deep foundation pit defect detection based on a small-sample depth long-tail self-learning algorithm, specifically includes the following steps: A deep foundation pit defect detection model is established, and the model loss function of the deep foundation pit defect detection model is calculated based on the labeled sample set obtained from the cruise. And according to the model loss function Update the parameters of the deep foundation pit defect detection model; The updated deep foundation pit defect detection model is used to detect foundation pit defects and obtain the defect bounding box of the foundation pit in the image. The defect rectangle (box) is used as a prompt word and input into the SAM model for refined extraction of crack defects, thereby obtaining the edge line of the crack.

2. The intelligent identification and high-precision measurement method for deep foundation pit defects according to claim 1, characterized in that: The use of GNSS For static observation stations, the calculation of the coordinates of the photogrammetric total station monitoring station using the corner resection method includes the following steps: According to the two presets GNSS Static observation station A , B With the total station of the photogrammetry reference monitoring station P Calculation of geometric positional relationship between the total station and the reference monitoring station P The coordinates of the points are determined, and the location of the total station monitoring station is calculated. P The mean error; Based on the aforementioned mean error, the total station monitoring station... P The point coordinates are corrected.

3. The intelligent identification and high-precision measurement method for deep foundation pit defects according to claim 2, characterized in that: According to the two presets GNSS Static observation station A , B With total station monitoring station P Calculation of geometric positional relationship between the total station and the reference monitoring station P The coordinates of the points are determined, and the location of the total station monitoring station is calculated. P The standard error specifically includes the following steps: Using a total station as a reference monitoring station P Measure its with GNSS Static observation station A The length of the side between S And its relationship with two GNSS Static observation station A , B The angle between β ; Using two GNSS Static observation station A , B Measure the distance between the two S 0. The direction angle between the two α AB as well as GNSS Static observation station A coordinates ( X A , Y A ), and calculate the coordinates of the reference monitoring station P of the total photogrammetric instrument ( XP , Yp The calculation formula is: （1）； （2）； （3）； （4）； （5）； in, for GNSS Static observation station B With the total station of the photogrammetry reference monitoring station P The connection and two GNSS Static observation station A , B The angle between the lines, for GNSS Static observation station A With the total station of the photogrammetry reference monitoring station P The connection and two GNSS Static observation station A , B The angle between the lines, a AP for GNSS Static observation station A With the total station of the photogrammetry reference monitoring station P The direction angle between them a AB For two GNSS Static observation station A , B The direction angle between; Substitute equation (3) into (4) and (5) and compare the total station reference monitoring station. P Measure its with GNSS Static observation station A The length of the side between S Photographic total station reference monitoring station P With two GNSS Static observation station A , B The angle between β Taking the differential, we get: （6）； Converting to the form of standard error yields: （7）； Then the reference monitoring station P of the total photogrammetric instrument ( XP , Yp The mean error of ) m p for: （8）； Simplifying and rearranging, we get: （9）； in, ρ It represents the radian constant in seconds.

4. The intelligent identification and high-precision measurement method for deep foundation pit defects according to claim 3, characterized in that: The process of acquiring calibration image information of a pre-set calibration plate using a total photogrammetric instrument, and then using the calibration image information and the coordinates of the monitoring station points of the total photogrammetric instrument to perform external parameter calibration of the total photogrammetric instrument, specifically includes the following steps: Using a total photostation, obtain calibration image information of a pre-set calibration plate, and then... SIFT The algorithm extracts the corner points of the chessboard grid in the calibration image information, obtains the image coordinates of the chessboard grid corner points, and matches them with the corresponding object points; The rotation vector between the image-side coordinates and the corresponding object-side coordinates is calculated using Zhang Zhengyou's calibration method. The rotation vector is converted into a rotation matrix using the Rodriguez transformation. R Complete the external parameter calibration of the total station.

5. The intelligent identification and high-precision measurement method for deep foundation pit defects according to claim 1, characterized in that: The model loss function of the deep foundation pit defect detection model is calculated based on the labeled sample set obtained from the cruise. Specifically, the steps include the following: The supervised loss function of the deep foundation pit defect detection model is calculated based on the labeled sample set obtained from the cruise. The calculation formula is: （9）； （10）； （11）； （12）； in, Indicates the first k One detection head, and They represent the first k The target probability score and detection box output by each detection head. and These represent the defect classification and bounding box loss functions, respectively. and These are the classification loss and bounding box regression loss for classifying foundation pit defects, respectively. , and These represent the input foundation pit image, foundation pit defect bounding box, and defect category, respectively. For the unlabeled sample set acquired during cruise operations First, PPYOLOE-F was used to detect foundation pit defects in the unlabeled data: （13）； in and They represent the first k The target probability scores and detection boxes output by each detection head are fused together to obtain pseudo-boxes. and its corresponding target probability score and categories : （13）； （14）； （15）； （16）； The statistical depth foundation pit defect detection model calculates the average probability score for each type of defect. and standard deviation vector And set a pseudo-box filtering threshold for each category. （17）； in, Indicates the first k Each detection head In the category of foundation pit c The filtering threshold above, and The tables show the categories of foundation pits. c Average probability score and standard deviation vector ; Calculate the unsupervised loss function : （18）； （19）； （20）； in, Indicates the first k Each detection head is based on a threshold. For category The pseudo-boundaries are filtered out, and the target probability score of the pseudo-boundary is... Greater than or equal to the threshold , When the value is 1, the pseudobox participates in the unsupervised loss function. The value is calculated and used in model training; otherwise, it is set to 0. Pseudoboxes do not participate in the calculation of the unsupervised loss function and do not participate in training. Based on the supervised loss function and unsupervised loss function Calculate the model loss function of a few-sample deep long-tail self-learning model. The calculation formula is: （21）； That is, the model loss function For supervised loss function and unsupervised loss function sum.

6. The intelligent identification and high-precision measurement method for deep foundation pit defects according to claim 1, characterized in that, The calculation of the offset angle between the defect point and the horizontal axis of the total station based on the pixel coordinates of the defect point specifically includes the following steps: Calculate the two-dimensional offset angle between the defect point and the principal axis of the camera based on the pixel coordinates of the defect point and the camera principal distance of the total station. α , β ): （22）； in,( x , y () represents the pixel coordinates of the defect point. f It is the camera image principal distance, ( x c , y c () represents the principal point coordinates; Calculate the offset angle between the defect point and the horizontal axis of the total station based on the calibration parameters of the total station. γ : （23）； in, R The rotation matrix used for calibrating a total station.

7. A system for intelligent identification and high-precision measurement of defects in deep foundation pits, characterized in that, It includes a calculation module, a calibration module, a training module, a recognition module, and a measurement module; The computing module is used to utilize GNSS For static observation stations, the coordinates of the total station monitoring points are calculated using the corner resection method. The calibration module is used to acquire calibration image information of a pre-set calibration plate using a photogrammetric total station, and to perform external parameter calibration on the photogrammetric total station using the calibration image information. The training module is used to continuously cruise the total station of the camera at the monitoring station to obtain a sample set of concrete support for deep foundation pits, and to train the deep foundation pit defect detection model online based on the small sample depth long tail self-learning algorithm. The recognition module is used to acquire target images online and use the trained deep pit defect detection model to identify the pixel coordinates of defect points in real time, and calculate the offset angle between the defect point and the horizontal axis of the total station based on the pixel coordinates of the defect point. The measurement module is used to control the total station to continuously perform three-dimensional coordinate measurements on the defect point according to the offset angle, and to obtain the parameter measurement information of the defect point. The training module utilizes the total station to continuously cruise at the total station monitoring station, acquiring a sample set of deep foundation pit concrete support, and specifically implements online defect identification training for the deep foundation pit defect detection model based on a small-sample depth long-tail self-learning algorithm as follows: A deep foundation pit defect detection model is established, and the model loss function of the deep foundation pit defect detection model is calculated based on the labeled sample set obtained from the cruise. And according to the model loss function Update the parameters of the deep foundation pit defect detection model; The updated deep foundation pit defect detection model is used to detect foundation pit defects and obtain the defect bounding box of the foundation pit in the image. The defect rectangle (box) is used as a prompt word and input into the SAM model for refined extraction of crack defects, thereby obtaining the edge line of the crack.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the intelligent identification and high-precision measurement method for deep foundation pit defects as described in any one of claims 1-6.

9. A device for intelligent identification and high-precision measurement of defects in deep foundation pits, characterized in that: It includes a communication interface, a memory, a communication bus, and a processor, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of the intelligent identification and high-precision measurement method for deep foundation pit defects as described in any one of claims 1 to 6.

Images