Non-contact monocular vision high-precision automatic settlement monitoring method

By using YOLOv10s model and digital image processing technology, combined with deep learning model, the problem of insufficient accuracy in target recognition, positioning and settlement value calculation of traditional settlement monitoring methods is solved, and high-precision settlement monitoring and early warning are achieved.

CN120252635APending Publication Date: 2025-07-04ANHUI ZHONGKEFENG INTELLIGENT TECHNOLOGY CO LTD +1
View PDF 0 Cites 4 Cited by

Patent Information

Application Number
CN202510541965.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-04-28
Publication Date
2025-07-04

AI Technical Summary

Technical Problem

Traditional settlement monitoring methods have limitations in cost, environmental adaptability and automation levels. The target identification, positioning and settlement value calculation accuracy are insufficient, especially in complex environments, and high-precision monitoring is difficult to achieve.

Method used

Target recognition is used by YOLOv10s model, target positioning is combined with digital image processing technology and least squares ellipse fitting algorithm, three-dimensional world coordinates are calculated through camera imaging principles, and the settlement value is predicted using deep learning models, reducing calculation overhead and improving monitoring accuracy.

Benefits of technology

It improves the target recognition and positioning accuracy, realizes high-precision settlement monitoring, and can monitor and predict future settlement trends in real time in complex environments, providing early warning for structural safety.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120252635A_ABST
    Figure CN120252635A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of settlement observation, in particular to a non-contact monocular vision high-precision automatic settlement monitoring method, which comprises the following steps: firstly, performing target monitoring by using a YOLOv10s model, improving the real-time reasoning speed and recognition precision of target recognition, and solving the target recognition problem under a complex background; secondly, a digital image processing technology and an edge-based least square ellipse fitting algorithm are combined to realize accurate positioning of the target; thirdly, deriving based on a camera imaging principle, and introducing an improved world coordinate calculation method to solve world coordinates corresponding to the pixel coordinates so as to obtain three-dimensional world coordinates; and finally, calculating a settlement value of the monitoring area by taking the first frame image as a reference, predicting a future settlement value by using a deep learning model, and making an early warning for structure safety in advance. According to the invention, the installation demand and the calculation overhead are reduced, and the settlement monitoring precision is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of settlement observation, and specifically to a non-contact monocular vision high-precision automatic settlement monitoring method. Background Technique

[0002] With the continuous acceleration of China's modernization process, the number of infrastructure construction projects is increasing day by day, and the safety of infrastructure such as roads, bridges, subways, and buildings has received more and more attention. However, traditional settlement monitoring methods have certain limitations in terms of cost, environmental adaptability, and automation level. For example, contact monitoring equipment often requires high installation and maintenance costs and is easily affected by external environmental factors, which affects the stability of monitoring.

[0003] In recent years, certain achievements have been made in using computer vision technology for settlement monitoring. The non-contact feature has effectively improved the monitoring efficiency and reduced the on-site layout requirements. However, existing solutions still have deficiencies in terms of monitoring accuracy in aspects such as target recognition, positioning, and settlement value calculation. In terms of target recognition, some solutions use methods such as template matching, which rely on the consistency of the target appearance and are easily affected by factors such as lighting and angle changes, resulting in unstable recognition accuracy. In addition, some object detection algorithms, such as SSD, still have room for improvement in terms of real-time inference speed and recognition accuracy of small targets in complex environments. Target positioning is another key link. The commonly used morphological processing and geometric feature matching methods in some solutions are difficult to achieve stable high-precision positioning in the case of complex backgrounds and diverse scenes. In the process of calculating the settlement value, accurately converting pixel coordinates into actual world coordinates is a technical difficulty. Some solutions use the direct linear transformation method, but in high-precision application scenarios, it is easily affected by noise, resulting in the amplification of errors and further affecting the overall monitoring effect.

[0004] Therefore, a non-contact monocular vision high-precision automatic settlement monitoring method is proposed. Summary of the Invention

[0005] The object of the present invention is to provide a non-contact monocular vision high-precision automatic settlement monitoring method, including: First, use the YOLOv10s model for target monitoring to improve the real-time inference speed and recognition accuracy of target recognition, and solve the problem of target recognition in complex backgrounds; Second, combine digital image processing technology and the least squares ellipse fitting algorithm based on edges to achieve precise positioning of the target; Third, based on the principle of camera imaging, an improved world coordinate calculation method is introduced to solve the world coordinates corresponding to pixel coordinates, and accurate three-dimensional world coordinates are obtained; Finally, taking the first frame image as a reference, the settlement value of the monitoring area is calculated, and a deep learning model is used to predict the future settlement value to give an early warning for structural safety. The present invention reduces the installation requirements and computational overhead and improves the accuracy of settlement monitoring.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A non-contact monocular vision high-precision automatic settlement monitoring method, including:

[0008] Step S1: Install a target in the monitoring area, take a picture of the target and obtain a target image, perform data augmentation on the target image to obtain a first data set; preprocess the first data set to obtain a second data set;

[0009] Step S2: Input the second data set into a basic model for training to obtain a trained YOLOv10s model; input the second data set into the YOLOv10s model for testing. If the confidence level of the YOLOv10s model for target detection is higher than a preset threshold, output a first image; otherwise, adjust the parameters of the YOLOv10s model and perform retraining;

[0010] Step S3: Use digital image processing technology to process the first image to obtain a second image; use the Canny edge detection algorithm to extract the edge features of the second image, and use the least squares ellipse fitting algorithm to fit the edge features to obtain the pixel coordinates of the center point of the target;

[0011] Step S4: Use the Zhang Zhengyou calibration method to obtain camera calibration, and calculate the projection distance between the camera optical center and the world coordinate point according to the camera calibration; establish a projection relationship between the world coordinates and the pixel coordinates according to the projection distance and the camera calibration to obtain the three-dimensional world coordinates of the center point of the target;

[0012] Step S5: Obtain the second images of consecutive frames, where the second images include a first frame image and subsequent frame images; use the three-dimensional world coordinates of the first frame image as the initial coordinates, and compare the three-dimensional world coordinates of the subsequent frame images with the initial coordinates to obtain the settlement value.

[0013] Furthermore, the process of data augmentation includes:

[0014] Obtain the target image, which includes targets at different distances, backgrounds, and angles;

[0015] Add Gaussian noise and salt-and-pepper noise to the target image to generate a noisy image;

[0016] Rotate the noisy image at different angles to generate multi-directional images;

[0017] Perform operations of brightness enhancement and brightness reduction on the multi-directional images to obtain the first dataset.

[0018] Furthermore, the preprocessing includes:

[0019] Use the labelimg software to label the first dataset in the yolo format to obtain the third dataset;

[0020] Divide the third dataset according to expert experience to obtain the second dataset;

[0021] Among them, the second dataset includes a training set, a validation set, and a test set. The training set is used to train the base model; the validation set is used to evaluate the base model after each round of training; the test set is used to test the YOLOv10s model.

[0022] Furthermore, the YOLOv10s model includes:

[0023] An input layer, which is used to adjust the size and normalize the images in the second dataset to obtain the third dataset;

[0024] A backbone network layer, which is used to extract the basic features of the third dataset;

[0025] A feature pyramid network layer, which is used to extract high-level features based on the basic features and assign weights to the high-level features by combining the adaptive feature fusion method to obtain a multi-scale feature map;

[0026] An output layer, which is used to output the target position information and the confidence of the third dataset.

[0027] Furthermore, the process of obtaining the pixel coordinates includes:

[0028] Perform grayscale conversion, image denoising, and binarization operations on the first image to obtain the second image;

[0029] Extract the edge information of the second image using the Canny edge detection algorithm to generate a binary edge map;

[0030] Extract the set of edge points in the shape of the target from the binary edge map, and use the adaptive threshold method to screen the set of edge points to obtain the edge features;

[0031] Use the least squares ellipse fitting algorithm and the adaptive weight adjustment method to perform ellipse fitting on the edge features to generate the elliptical shape of the target;

[0032] Locate the center point of the target using the moment method based on the elliptical shape to obtain the pixel coordinates.

[0033] Further, the calculation process of the three-dimensional world coordinates includes:

[0034] Obtain the camera calibration and calculate the projection distance. The camera calibration includes the internal parameter matrix, the rotation matrix, and the translation vector;

[0035] Establish the projection relationship between the world coordinates and the pixel coordinates according to the camera imaging principle to obtain a projection function, which is expressed as:

[0036] s*P C =M*(R*P w +T);

[0037] where s is the projection distance; P C =(u,v,1) is the pixel coordinate, and u and v are the normalized pixel coordinate values on the image plane; M is the internal parameter matrix; R is the rotation matrix; P w =(X,Y,Z) is the world coordinate, and X, Y, and Z are the coordinate values of the three-dimensional world respectively; T is the translation vector;

[0038] Substitute the projection distance and the camera calibration into the projection function to obtain the three-dimensional world coordinates.

[0039] Further, the calculation process of the three-dimensional world coordinates also includes:

[0040] If there is more than one target in the target image, use the YOLOv10s model to identify each target one by one, and use geometric correction and depth estimation to correct the error of each three-dimensional world coordinate.

[0041] Further, the calculation process of the projection distance includes:

[0042] Set the vertical coordinate value of the world coordinate to 0;

[0043] The projection distance is calculated based on the camera calibration and the pixel coordinates, expressed as:

[0044] s = (Z + M2[2]) / M1[2];

[0045] where s is the projection distance; Z is the vertical coordinate value; M2[2] is the third component of the translation vector T projected into the camera coordinate system; M1[2] is the depth information of the pixel point P C in the camera coordinate system after conversion.

[0046] Furthermore, the process of obtaining the settlement value includes:

[0047] Obtain the first frame of image, and use the three-dimensional world coordinates of the first frame of image as the initial coordinates;

[0048] Obtain the subsequent frames of images, and use the three-dimensional world coordinates of the subsequent frames of images as the subsequent coordinates;

[0049] Subtract the subsequent coordinates in the vertical direction from the initial coordinates to obtain the settlement value;

[0050] Correspond each frame timestamp with the settlement value to obtain settlement change data;

[0051] Use the trained deep learning model to predict the future settlement value based on the settlement change data;

[0052] Divide the future settlement value into geological uplift and geological subsidence, and divide the severity level according to the severity of the geological uplift and the geological subsidence. The severity level includes the first level, the second level, and the third level;

[0053] If the future settlement value is the second level, a warning is issued; if the future settlement value is the third level, a warning is issued and manual intervention is reminded.

[0054] Furthermore, the process of obtaining the future settlement value includes:

[0055] Use the sliding window method to input the settlement change data into the second basic model for training to obtain the trained deep learning model;

[0056] where, if the length of the sliding window is w, the input data of the second basic model is w settlement change data;

[0057] Input the latest settlement change data into the deep learning model for prediction to obtain the future settlement value, and the future settlement value is the settlement value at the next moment.

[0058] Compared with the prior art, the beneficial effects of the present invention are:

[0059] 1. In terms of target recognition, when traditional target recognition methods deal with complex background images, the detection accuracy is often low due to background interference. By using the YOLOv10s model, not only the recognition accuracy of targets at different positions and sizes is improved, but also the computational cost is effectively reduced. In addition, the feature pyramid network layer and the adaptive feature fusion method are introduced to achieve the effective allocation of features at different scales, providing a data basis for subsequent settlement monitoring and thus improving the accuracy of settlement monitoring.

[0060] 2. In terms of target positioning, the positioning accuracy of traditional positioning methods needs to be further improved under complex imaging conditions. By processing the target image through digital image processing technology, the quality of the image is improved. In addition, the least squares ellipse fitting algorithm based on Canny edges is used, which can accurately fit the contour of the target and obtain the pixel coordinates of the target center point, not only effectively eliminating noise interference, but also improving the fitting accuracy of the target shape and the stability of positioning under non-ideal imaging conditions.

[0061] 3. During the process of calculating the settlement value, traditional settlement calculation methods have problems with insufficient accuracy in dynamic monitoring. By converting the pixel coordinates of the target into three-dimensional world coordinates, the calculation of the settlement value of the target is realized, which not only improves the accuracy of settlement monitoring, but also can dynamically monitor the settlement changes. In addition, combining with a deep learning model to predict the trend of settlement change data can meet the settlement monitoring requirements in different environments and provide a scientific basis for early warning and risk prevention and control. Description of the Drawings

[0062] Figure 1 Schematic diagram of the experimental site layout of the embodiment of the present invention;

[0063] Figure 2 Schematic diagram of the technical route process proposed in the embodiment of the present invention;

[0064] Figure 3 Schematic diagram of the target of the embodiment of the present invention;

[0065] Figure 4 Schematic diagram of the target detection effect of the embodiment of the present invention;

[0066] Figure 5 Schematic diagram of the target image processing effect of the embodiment of the present invention. Detailed Embodiments

[0067] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0068] With the continuous acceleration of China's modernization process, the demand for infrastructure such as highways, bridges, subways, and buildings is increasing day by day. The number of infrastructure construction projects is increasing continuously, and many new requirements are put forward for settlement monitoring. However, the traditional settlement monitoring methods have high costs, often requiring expensive professional equipment or complex installation and debugging processes. Secondly, the monitoring results are easily affected by external environmental factors such as light, climatic conditions, and surrounding noise, resulting in insufficient stability and accuracy of the data. In addition, the automation level of the traditional methods is relatively low, and it is difficult to achieve large-scale and long-term real-time monitoring. In the existing solutions, although some technologies have tried to improve the automation and detection accuracy, with the development of computer vision technology, some researchers monitor the settlement amount of the target through computer vision technology to complete the settlement monitoring of the monitoring area. However, the robustness under complex environments and dynamic conditions is still insufficient, affecting the accuracy of the monitoring results. Therefore, a non-contact monocular vision high-precision automatic settlement monitoring method is needed to reduce the computational overhead and improve the accuracy of settlement monitoring.

[0069] Please refer to Figures 1 to 5 , the present invention provides a non-contact monocular vision high-precision automatic settlement monitoring method, and the technical solution is as follows:

[0070] Embodiment 1

[0071] As Figure 1As shown in the figure, the experiment of this embodiment is carried out on an outdoor experimental site. The experimental equipment includes a Trimble DINI electronic level, a leveling staff, a target, and a fixed-focus camera with a focal length of 4 mm. The leveling staff is placed vertically beside the target and in the same plane as the target to ensure that the leveling staff and the target can be photographed by the camera at the same time, providing a reference point for position changes; the electronic level is placed at a certain horizontal distance from the target and the leveling staff to obtain the point position data at this time; the camera is fixed directly in front of the target, and the target image is collected by the camera at a frequency of 1 image every 15 seconds. In addition, in this embodiment, the software environment is Windows 11, the platform used is a laptop computer, the configuration is AMD Ryzen 7 7745HX@3.60 GHz, the operating memory is 16G, NVIDIA Ge-Force RTX 4060 Laptop GPU, and the video memory is 8GB. The codes used in the experiment are all written in Python. The purpose is to verify that the present invention can effectively reduce the computational overhead and improve the accuracy of settlement monitoring.

[0072] As Figure 2 shown, a non-contact monocular vision high-precision automatic settlement monitoring method includes:

[0073] Step S1: Install a target in the monitoring area, photograph the target and obtain the target image, perform data enhancement on the target image to obtain a first data set; perform preprocessing on the first data set to obtain a second data set.

[0074] Further, the process of data enhancement includes:

[0075] Obtain the target image, and the target image includes targets at different distances, backgrounds, and angles;

[0076] Specifically, in this embodiment, as Figure 3 shown, print the target picture and paste it on a PVC board, fix it with a tripod, and use it as the main target for measurement. Use the camera to take 271 target pictures at different distances, backgrounds, and angles. Settlement monitoring has high requirements for processing speed. Encoded markers have complex structures, high imaging requirements, and require multi-point matching, while non-encoded markers have simple structures, low imaging requirements, and low resolution requirements. Therefore, non-encoded markers are selected as the targets.

[0077] Add Gaussian noise and salt-and-pepper noise to the target image to generate a noisy image;

[0078] Specifically, noise can be added to the image to simulate the effects of the camera shooting in different environments, so that the model can still recognize the target in the noisy image, thereby improving the recognition stability of the model under noise conditions;

[0079] Rotate the noisy image at different angles, rotating clockwise and counterclockwise by multiple angles respectively, so as to generate multi-directional images, simulating the recognition of the target image at different tilt angles;

[0080] Perform operations of brightness enhancement and brightness reduction on the multi-directional images to obtain the first data set, improving the robustness of the model under natural light change conditions;

[0081] Among them, the first data set includes 1848 augmented target pictures. By generating target images with various illuminations and directions, the sample range of the data set is expanded, enabling the recognition model to more comprehensively learn the feature representations of the target at different angles and under different illumination conditions during training, reducing the situations of misrecognition and missed recognition, and thus improving the recognition accuracy of the target. For example, under dim or backlight conditions, if the enhanced low-light images have been seen during the training process, the recognition model can also effectively recognize the target.

[0082] Furthermore, the preprocessing includes:

[0083] Use the labelimg software to label the first data set in the yolo format to obtain the third data set;

[0084] Specifically, all targets in each target image will be boxed and assigned the same label "target", ensuring that the model can recognize and learn the features of the target;

[0085] Divide the third data set according to expert experience to obtain the second data set;

[0086] Among them, the second data set includes a training set, a validation set and a test set. The training set is used to train the basic model; the validation set is used to evaluate the basic model after each round of training; the test set is used to test the YOLOv10s model.

[0087] Among them, the training set, the validation set and the test set are randomly allocated according to the ratio of 7:1:2. This data division method can not only effectively support the training of the model, but also ensure that representative data can be obtained during the validation and test processes, thus improving the accuracy and reliability of the YOLOv10s model in practical applications.

[0088] Step S2: Input the second data set into the basic model for training to obtain a trained YOLOv10s model; input the second data set into the YOLOv10s model for testing. If the confidence level of the YOLOv10s model for target detection is higher than the preset threshold, output the first image, otherwise adjust the parameters of the YOLOv10s model and perform retraining;

[0089] Among them, the preset threshold is set to 0.9.

[0090] Specifically, YOLO (You Only Look Once), as an efficient object detection algorithm, performs excellently in terms of real-time performance and accuracy. The latest version of YOLO, YOLOv10, introduces a consistent dual-task method, achieving training without non-maximum suppression, thereby reducing the inference latency. In addition, compared with other versions, YOLOv10s is easier to tune and optimize, has a faster inference speed, and is more suitable for settlement monitoring. Therefore, YOLOv10s is selected as the target recognition model.

[0091] Furthermore, the YOLOv10s model includes:

[0092] An input layer for resizing and normalizing the images of the second dataset to generate an image set suitable for model processing, namely the third dataset;

[0093] A backbone network layer for extracting the basic features of the third dataset; among them, the backbone network layer uses a convolutional neural network to gradually extract basic features through multi-layer convolution and pooling operations, such as edge and texture information;

[0094] A feature pyramid network layer for extracting high-level features based on the basic features and assigning weights to the high-level features by combining an adaptive feature fusion method to obtain a multi-scale feature map;

[0095] Specifically, the feature pyramid network layer enables the model to obtain the details and global information of the target from low to high resolution by extracting different scale features of the target image layer by layer. For example, larger targets are detected through low-resolution feature maps, while high-resolution feature maps are used to identify fine target details, ensuring that each target can be accurately located. In addition, the adaptive feature fusion method further analyzes the importance of each scale feature map, adds or fuses the feature maps, realizes the multi-level feature integration of the target, and improves the detection accuracy of the model. For example, when the target is far away or the lighting conditions are not ideal, the deep features will obtain higher weights, ensuring that the target is still correctly recognized at a lower resolution.

[0096] An output layer for outputting the target position information and the confidence of the third dataset.

[0097] Specifically, in the trained YOLOv10s model, the training parameters are as follows: the learning rate is 0.01, the batch size is 6, and the image size is 1440 pixels. To prove the effectiveness of the YOLOv10s model, other models were used to detect the target. As shown in Table 1, the accuracy (mAP) represents the detection consistency and confidence level in multiple frames of images, and the speed (FPS) represents the number of frames of images processed per second. It can be seen that the YOLOv10s model performs best in terms of both accuracy and speed, is suitable for the target detection task under complex backgrounds, and ensures the detection accuracy and real-time performance in settlement monitoring.

[0098] Table 1 Comparison of target detection accuracy and speed of different detection models

[0099] model template matching SSD Faster R-CNN Mask R-CNN YOLOV5s YOLOv10s Accuracy (mAP) 0.65 0.72 0.82 0.84 0.88 0.91 Speed (FPS) 10 15 5 3 20 25

[0100] The trained YOLOv10s model was tested for target detection, and the test results are as Figure 4 shown. It can be seen that in different scenarios, the confidence level of the trained YOLOv10s model for target detection is above 0.9, indicating that the trained model can meet the target detection requirements in the settlement monitoring task.

[0101] Step S3: Process the first image using digital image processing technology to obtain a second image; extract the edge features of the second image using the Canny edge detection algorithm, and fit the edge features using the least squares ellipse fitting algorithm to obtain the pixel coordinates of the target center point.

[0102] Among them, digital image processing technology includes operations such as grayscale conversion, image denoising, and binarization.

[0103] Specifically, in this embodiment, as Figure 5 shown, Figure 5 The upper part shows the detection results using the YOLOv10s model. The current confidence score of this target detection is 0.93, indicating that the detection accuracy meets the requirements. Figure 5 The lower part shows the processing effect of the target image. The entire processing process includes sequentially performing grayscale conversion, binarization, denoising, edge detection, and center positioning on the original target image, and finally the pixel coordinates of the target center point can be obtained.

[0104] Furthermore, the process of obtaining the pixel coordinates includes:

[0105] Perform grayscale conversion, image denoising, and binarization operations on the first image to obtain the second image;

[0106] Specifically, first, the first image is converted into a grayscale image to reduce the complexity of the image data and make the edge features more obvious. Then, median filtering is used to reduce the interference of random noise on edge detection. The window size is set to 3×3 to ensure the smoothness of the image while retaining the edge information. Finally, the Otsu adaptive threshold method is used to divide the pixels into target and background areas to generate a clear binary image.

[0107] Extract edge information of the second image using the Canny edge detection algorithm to generate a binary edge map;

[0108] Specifically, a dual threshold strategy is used for edge filtering to retain stronger edges and suppress weak edges, where the low threshold and high threshold are set to 0.1 and 0.3.

[0109] Further extracting a target-shaped edge point set from the binary edge image, and screening the edge point set using an adaptive threshold method to obtain the edge feature;

[0110] Using the least squares ellipse fitting algorithm and the adaptive weight adjustment method to perform ellipse fitting on the edge features to generate an ellipse shape of the target;

[0111] The target center point is located using the moment method according to the ellipse shape to obtain the pixel coordinates.

[0112] Specifically, during the fitting process of edge features, the fitting errors of all edge points are calculated, and edge points with large deviations are eliminated to improve the accuracy of fitting. After the fitting is completed, all parameters of the ellipse are obtained, and the center point of the ellipse is extracted from them as the positioning information of the target.

[0113] Using digital image processing technology to process the first image can reduce the complexity of the image and enhance the recognizability of edge features. The Canny edge detection algorithm combined with the dual threshold strategy can reduce the influence of pseudo edges and ensure that the extracted edge information is more accurate. The least squares method is used to obtain the positioning information of the target, which provides a precise target position for subsequent settlement monitoring, so that the position changes of the target can be fed back in time.

[0114] Step S4: using the Zhang Zhengyou calibration method to obtain camera calibration, and calculating the projection distance between the camera optical center and the world coordinate point according to the camera calibration; establishing the projection relationship between the world coordinate and the pixel coordinate according to the projection distance and the camera calibration, and obtaining the three-dimensional world coordinate of the target center point;

[0115] Furthermore, the calculation process of the three-dimensional world coordinates includes:

[0116] Obtain the camera calibration and calculate the projection distance. The camera calibration includes an intrinsic matrix, a rotation matrix, and a translation vector;

[0117] Among them, the intrinsic matrix is used to describe the internal characteristics of the camera, such as the focal length and the position of the optical center; the rotation matrix is used to describe the rotation relationship between the camera coordinate system and the world coordinate system, and is a 3×3 orthogonal matrix; the translation vector is used to describe the position of the camera in the world coordinate system;

[0118] Establish the projection relationship between the world coordinates and the pixel coordinates according to the camera imaging principle to obtain a projection function, which is expressed as:

[0119] s*P C =M*(R*P w +T);

[0120] Among them, s is the projection distance; P C =(u,v,1) is the pixel coordinate, where u and v are the normalized pixel coordinate values on the image plane; M is the intrinsic matrix; R is the rotation matrix; P w =(X,Y,Z) is the world coordinate, where X, Y, and Z are the coordinate values in the three-dimensional world respectively; T is the translation vector;

[0121] Transform the projection function to solve for P w , which is expressed as:

[0122] R -1 *[M -1 *(s*P C )-T]=P w ;

[0123] Among them, R -1 is the inverse matrix of the rotation matrix, and M -1 is the inverse matrix of the intrinsic matrix;

[0124] Substitute the projection distance and the camera calibration into the projection function to obtain the three-dimensional world coordinates.

[0125] Specifically, assume that the pixel coordinate of a certain target in the image is P C =(320,240,1), and the intrinsic matrix of the camera is:

[0126]

[0127] Among them, the focal length is 1000 pixels, and the optical center coordinates are (400,300);

[0128] Assume that the target has a 45-degree rotation relative to the camera, which is expressed as:

[0129]

[0130] If the camera is 2 meters away from the target, the translation vector is:

[0131]

[0132] Assuming the projection distance s is 1.5 meters, substituting this information into the projection function, P can be calculated. w It is (0.1, -0.01, 3.5). By combining the YOLO detection model to obtain the pixel coordinates of the target and using the camera parameters to convert the pixel coordinates into three-dimensional world coordinates, high-precision positioning of the target can be provided, realizing an accurate mapping from the image to the real physical space. In addition, this calculation method can effectively reduce the amount of calculation, be able to obtain a quick response in practice, and thus reduce the calculation burden.

[0133] Furthermore, the calculation process of the three-dimensional world coordinates further includes:

[0134] If there is more than one target in the target image, use the YOLOv10s model to identify each target one by one, and use geometric correction and depth estimation to correct the error of each three-dimensional world coordinate.

[0135] Specifically, when multiple targets are at different depth positions, calculation errors may occur. For example, the pixel coordinates of target A and target B are P Ca =(320, 240, 1) and P Cb =(500, 300, 1) respectively. Using geometric correction to correct the pixel coordinates, the pixel coordinates of target A and target B become P Ca =(320, 238, 1) and P Cb =(498, 298, 1) respectively, making the corrected coordinates closer to the ideal plane coordinates, reducing the deviation introduced by lens distortion, and thus improving the accuracy of settlement monitoring. In addition, the calculated three-dimensional world coordinates can be further corrected using depth estimation according to the known relative relationship of the targets to ensure the accuracy of the three-dimensional coordinates. For example, assuming the three-dimensional world coordinates of target A and target B are P wa =(0.1, -0.01, 3.5) and P wb =(0.4, 0.02, 4.8) respectively. Since the actual distance between target A and target B is 0.5 meters, while the calculated distance is 1.33 meters, the depth Z values of A and B need to be adjusted until the spatial distance is 0.5 meters. The finally corrected three-dimensional coordinates are P wa =(0.1, -0.01, 2.5) and P wb =(0.4, 0.02, 3.8).

[0136] Further, the calculation process of the projection distance includes:

[0137] Continuing to transform the projection function to solve for s, expressed as:

[0138]

[0139] where s is the projection distance; P C =(u, v, 1) is the pixel coordinate, and u and v are the normalized pixel coordinate values on the image plane; M is the internal parameter matrix; R is the rotation matrix; P w =(X, Y, Z) is the world coordinate, and X, Y, and Z are the coordinate values in the three-dimensional world respectively; T is the translation vector;

[0140] Set the vertical coordinate value Z of the world coordinate to 0;

[0141] Calculate the projection distance according to the camera calibration and the pixel coordinate, expressed as:

[0142] s = (Z + M2[2]) / M1[2];

[0143] where s is the projection distance; Z is the vertical coordinate value; M2[2] is the third component of the translation vector T projected into the camera coordinate system, that is, the third element value of R -1 *T (column vector); M1[2] is the depth information after conversion of the pixel point P C in the camera coordinate system, that is, the third element value of R -1 *M -1 *P C (column vector).

[0144] Specifically, in monocular camera calibration, there are various methods for estimating the internal and external parameters of the camera, such as Zhang Zhengyou's method, direct linear transformation, self-calibration method based on quadratic surfaces, and self-calibration method based on Kruppa equations. The direct linear transformation method requires precise three-dimensional calibration points and complex linear algebra calculations. In contrast, Zhang Zhengyou's method only requires the two-dimensional information of the planar calibration board, obtains the calibration points through corner detection of the checkerboard template, is convenient for actual operation, and can obtain stable calibration results in different poses. In addition, the self-calibration method based on quadratic surfaces and the self-calibration method based on Kruppa equations usually rely on motion assumptions or system-built-in constraint conditions, and the calibration process is complex. However, Zhang Zhengyou's method can calibrate the monocular camera without additional motion conditions and has better robustness to image noise. Therefore, it can calculate the camera parameters and the distance of the object more accurately, thereby improving the accuracy of settlement monitoring.

[0145] Specifically, to measure the monitoring accuracy of the proposed algorithm, by changing the distance between the camera and the target, the relationship between the monitoring accuracy of the present invention and the distance was studied, and two evaluation indicators were selected: the absolute error between the visually calculated value and the value measured by the electronic level, and the mean absolute error (MAE).

[0146] As shown in Table 2, in the entire experimental data, when the camera is 5 meters away from the target, the maximum absolute error is 2.784 mm, the minimum absolute error is 0.246 mm, and the mean absolute error is 1.151 mm. As shown in Table 3, when the camera is 10 meters away from the target, the maximum absolute error is 4.071 mm, the minimum absolute error is 0.42 mm, and the mean absolute error is 1.543 mm. It can be seen from this that this method performs well in terms of feasibility and accuracy. In addition, the monitoring accuracy is affected by the distance between the target and the camera, and the farther the distance, the greater the corresponding error.

[0147] Table 2 Experimental results when the camera is 5 meters away from the target

[0148]

[0149] Table 3 Experimental results when the camera is 10 meters away from the target

[0150]

[0151] Step S5: Obtain the second images of consecutive frames, where the second images include the first frame image and subsequent frame images; use the three-dimensional world coordinates of the first frame image as the initial coordinates, and compare the three-dimensional world coordinates of the subsequent frame images with the initial coordinates to obtain the settlement value.

[0152] Furthermore, the process of obtaining the settlement value includes:

[0153] Obtain the first frame image, denoted as I0;

[0154] Use the projection function to calculate the three-dimensional world coordinates, and use the three-dimensional world coordinates of the first frame image as the initial coordinates, denoted as P w0 =(x0,y0,z0);

[0155] After 15 s, obtain the subsequent frame images, denoted as I n ;

[0156] Similarly, use the projection function to calculate the three-dimensional world coordinates, and use the three-dimensional world coordinates of the subsequent frame images as the subsequent coordinates, denoted as P wi =(x i ,y i ,z i );

[0157] Subtract the initial coordinate in the vertical direction from the subsequent coordinate to obtain the settlement value, expressed as:

[0158] c i = P wi [z] - P w0 [z];

[0159] where c i is the settlement value, P wi [z] is the vertical component of the subsequent coordinate, and P w0 [z] is the vertical component of the initial coordinate;

[0160] Correspond each frame timestamp with the settlement value to obtain settlement change data, expressed as:

[0161] {(t0, c0), (t1, c1), (t2, c2), …, (t n , c n )};

[0162] where t0 to t n are timestamps, and c0 to c n are settlement values;

[0163] Predict the future settlement value according to the settlement change data using a trained deep learning model.

[0164] Among them, the deep learning model can use models such as the LSTM model and the GRU model.

[0165] Divide the future settlement value into geological uplift and geological subsidence, and divide the severity level according to the severity of the geological uplift and the geological subsidence. The severity level includes the first level, the second level, and the third level;

[0166] If the future settlement value is the second level, a warning is issued; if the future settlement value is the third level, a warning is issued and manual intervention is reminded.

[0167] Among them, geological subsidence and geological uplift are represented by negative values and positive values respectively. After taking the absolute value of the values, the severity level can be divided into [0, 10), [10, 20), and [20, +∞).

[0168] By recording the timestamp and the corresponding settlement value frame by frame, the dynamic change of the structure settlement can be monitored in real time, which helps to judge the safety state of the settlement structure in time, provides reliable data support for engineering design and subsequent construction management, and realizes the accurate monitoring of the settlement value. In addition, with the prediction function of the deep learning model, the future settlement trend can be predicted according to the settlement change data, realizing early warning.

[0169] Furthermore, the process of obtaining the future settlement value includes:

[0170] Using a sliding window method, input the settlement change data into the second basic model for training to obtain the trained deep learning model; wherein, if the length of the sliding window is w, the input data of the second basic model is w settlement change data for model training and optimization;

[0171] Input the latest settlement change data into the deep learning model for prediction to obtain the future settlement value, which is the settlement value at the next moment.

[0172] Specifically, conduct an experiment on soft soil settlement to monitor the soil settlement under different loads. In the case of load A, the soft soil settlement change data is shown in Table 4, and each data point contains a timestamp t i and the corresponding settlement value c i , select the sliding window length w to be 5, then the data corresponding to input window 1 is [0.0, 0.2, 0.5, 0.9, 1.4], and each window slides 15 seconds to cover new data points. The second basic model selects the LSTM model, input the settlement change data of the first 7 days into the LSTM model for training, and the mean square error of the LSTM model is 0.12. Input the data of the 8th day into the LSTM model for testing, the model output prediction result is 6.4 mm, and the actual observed value is 6.5 mm, and the prediction error is only 0.1 mm, indicating that the model has a high prediction accuracy. Through the feature extraction of the time series, the model can accurately predict the future settlement trend, which helps to prevent potential risks in advance and facilitates maintenance personnel to take protective measures in a timely manner.

[0173] Table 4 Partial Settlement Change Data

[0174] Timestamp t (s) Settlement value c (mm) 0 0.0 15 0.2 30 0.5 45 0.9 60 1.4 75 2.0

[0175] Example 2

[0176] The present invention only requires a monocular camera and a target, which has a lower cost compared to traditional multi-sensor solutions. It obtains settlement information through a non-contact method, avoiding errors introduced by contact detection and wear and tear of instruments. By using image enhancement technology and the YOLOv10s model to detect the target, the recognition accuracy of the target is ensured. The Canny edge detection algorithm and the least squares ellipse fitting algorithm are used to extract the pixel coordinates of the target center point, ensuring the accuracy of edge recognition in complex backgrounds and helping to accurately determine the target position in the image. Through projection distance and camera calibration, the pixel coordinates are accurately matched with the three-dimensional world coordinates, capable of adapting to perspective changes in different monitoring scenarios. Finally, a deep learning model is used to predict future settlement values, which can accurately predict future settlement trends and help prevent potential risks in advance.

[0177] To further prove the effectiveness of the present invention, in this embodiment, a certain company conducts settlement monitoring on a certain building under its jurisdiction and uses existing monitoring data to predict future settlement values, aiming to improve the accuracy of settlement monitoring.

[0178] A non-contact monocular vision high-precision automatic settlement monitoring method includes:

[0179] Step S1: Install a target in the monitoring area, photograph the target and obtain a target image, perform data enhancement on the target image to obtain a first data set; preprocess the first data set to obtain a second data set;

[0180] Among them, multiple targets are installed at key settlement monitoring positions of the building, such as building foundations, load-bearing wall corners, and support columns. The surface of the target adopts an easily recognizable pattern, such as black and white rings.

[0181] Step S2: Input the second data set into a basic model for training to obtain a trained YOLOv10s model; input the second data set into the YOLOv10s model for testing. If the confidence level of the YOLOv10s model in detecting the target is higher than a preset threshold, output a first image; otherwise, adjust the parameters of the YOLOv10s model and perform retraining;

[0182] Among them, the preset threshold is set to 0.85.

[0183] Step S3: Use digital image processing technology to process the first image to obtain a second image; use the Canny edge detection algorithm to extract the edge features of the second image, and use the least squares ellipse fitting algorithm to fit the edge features to obtain the pixel coordinates of the target center point;

[0184] Step S4: Obtain camera calibration using Zhang Zhengyou calibration method, and calculate the projection distance between the camera optical center and the world coordinate point according to the camera calibration; establish the projection relationship between the world coordinate and the pixel coordinate according to the projection distance and the camera calibration to obtain the three-dimensional world coordinate of the center point of the target;

[0185] Step S5: Obtain the second images of consecutive frames, where the second images include the first frame image and subsequent frame images; use the three-dimensional world coordinate of the first frame image as the initial coordinate, and compare the three-dimensional world coordinate of the subsequent frame images with the initial coordinate to obtain the settlement value.

[0186] Further, the process of data augmentation includes:

[0187] Continuously photograph the building monitoring points using a high-resolution camera to obtain multiple frames of the target images, where the target images include targets at different distances, backgrounds, and angles;

[0188] Add Gaussian noise and salt-and-pepper noise to the target images to generate noisy images;

[0189] Rotate the noisy images at different angles to generate multi-directional images;

[0190] Perform operations of brightness enhancement and brightness reduction on the multi-directional images to obtain the first data set.

[0191] Further, the preprocessing includes:

[0192] Use labelimg software to label the first data set in yolo format to obtain the third data set;

[0193] Divide the third data set according to expert experience to obtain the second data set;

[0194] Among them, the second data set includes a training set, a validation set, and a test set. The training set is used to train the basic model; the validation set is used to evaluate the basic model after each round of training; the test set is used to test the YOLOv10s model.

[0195] Among them, the ratio of the training set, the validation set, and the test set is 8:1:1.

[0196] Further, the YOLOv10s model includes:

[0197] An input layer, which is used to adjust the size and normalize the images of the second data set to obtain the third data set;

[0198] A backbone network layer, which is used to extract the basic features of the third data set;

[0199] The feature pyramid network layer is used to extract high-level features based on the basic features, and combine the adaptive feature fusion method to assign weights to the high-level features to obtain a multi-scale feature map;

[0200] The output layer is used to output the target position information and the confidence of the third data set.

[0201] Further, the process of obtaining the pixel coordinates includes:

[0202] Gray-scale the first image, perform image denoising and binarization operations to obtain the second image;

[0203] Use the Canny edge detection algorithm to extract the edge information of the second image and generate a binary edge map;

[0204] Extract the set of edge points of the target shape from the binary edge map, and use the adaptive threshold method to screen the set of edge points to obtain the edge features;

[0205] Use the least squares ellipse fitting algorithm and the adaptive weight adjustment method to perform ellipse fitting on the edge features to generate the elliptical shape of the target;

[0206] Locate the center point of the target using the moment method according to the elliptical shape to obtain the pixel coordinates.

[0207] Further, the process of calculating the three-dimensional world coordinates includes:

[0208] Obtain the camera calibration and calculate the projection distance. The camera calibration includes the internal parameter matrix, the rotation matrix, and the translation vector;

[0209] Establish the projection relationship between the world coordinates and the pixel coordinates according to the camera imaging principle to obtain the projection function, which is expressed as:

[0210] s*P C =M*(R*P w +T);

[0211] where s is the projection distance; P C =(u,v,1) is the pixel coordinate, and u and v are the normalized pixel coordinate values on the image plane; M is the internal parameter matrix; R is the rotation matrix; P w =(X,Y,Z) is the world coordinate, and X, Y, and Z are the coordinate values of the three-dimensional world respectively; T is the translation vector;

[0212] Substitute the projection distance and the camera calibration into the projection function to obtain the three-dimensional world coordinates.

[0213] Further, the calculation process of the three-dimensional world coordinates further includes:

[0214] If there is more than one target in the target image, use the YOLOv10s model to identify each target one by one, and use geometric correction and depth estimation to correct the error of each three-dimensional world coordinate.

[0215] Further, the calculation process of the projection distance includes:

[0216] Set the vertical coordinate value of the world coordinates to 0;

[0217] Calculate the projection distance according to the camera calibration and the pixel coordinates, expressed as:

[0218] s = (Z + M2[2]) / M1[2];

[0219] where s is the projection distance; Z is the vertical coordinate value; M2[2] is the third component of the translation vector T projected into the camera coordinate system; M1[2] is the depth information after conversion of the pixel point P in the camera coordinate system. C Converted depth information.

[0220] Further, the process of obtaining the settlement value includes:

[0221] Obtain the first frame image, and use the three-dimensional world coordinates of the first frame image as the initial coordinates;

[0222] Obtain the subsequent frame images, and use the three-dimensional world coordinates of the subsequent frame images as the subsequent coordinates;

[0223] Subtract the subsequent coordinates in the vertical direction from the initial coordinates to obtain the settlement value;

[0224] Correspond each frame timestamp with the settlement value to obtain settlement change data;

[0225] Use the trained deep learning model to predict the future settlement value according to the settlement change data.

[0226] Divide the future settlement value into geological uplift and geological subsidence, and divide the severity level according to the severity of the geological uplift and the geological subsidence. The severity level includes the first level, the second level, and the third level;

[0227] If the future settlement value is the second level, issue a warning; if the future settlement value is the third level, issue a warning and remind for manual intervention.

[0228] Among them, geological subsidence and geological uplift are represented by negative and positive values respectively, and the severity levels can be divided into [0, 10), [10, 20) and [20, +∞), and the divided severity levels are all in millimeters.

[0229] Further, the process of obtaining the future settlement value includes:

[0230] Using the sliding window method to input the settlement change data into the second basic model for training to obtain the trained deep learning model;

[0231] Among them, if the length of the sliding window is w, the input data of the second basic model is w settlement change data;

[0232] Input the latest settlement change data into the deep learning model for prediction to obtain the future settlement value, and the future settlement value is the settlement value at the next moment.

[0233] Specifically, in the actual monitoring process, the system collects building settlement data every 1 hour. After 6 months of monitoring, using the collected data as shown in Table 5, through preliminary analysis of these data, it can be seen that the change in the settlement value gradually increases and is accompanied by certain fluctuations. Use the GRU model to predict the settlement value, and predict the settlement value at the next time step. During the predicted one-year period, the settlement value shows a gradually increasing trend in the next year, and the monthly settlement increment increases, until it gradually levels off in June of the following year. When the settlement value reaches the third level, such as 21 millimeters, structural reinforcement or other preventive maintenance can be carried out.

[0234] Table 5 Partial settlement change data

[0235]

[0236] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A non-contact monocular vision high-precision automatic settlement monitoring method, characterized in that Including: Step S1: Install a target in the monitoring area, take a picture of the target and obtain a target image, perform data augmentation on the target image to obtain a first data set; Preprocess the first data set to obtain a second data set; Step S2: Input the second data set into a basic model for training to obtain a trained YOLOv10s model; input the second data set into the YOLOv10s model for testing. If the confidence level of the YOLOv10s model in target detection is higher than a preset threshold, output a first image; otherwise, adjust the parameters of the YOLOv10s model and perform retraining; Step S3: Use digital image processing technology to process the first image to obtain a second image; use the Canny edge detection algorithm to extract the edge features of the second image, and use the least squares ellipse fitting algorithm to fit the edge features to obtain the pixel coordinates of the target center point; Step S4: Use the Zhang-Zhengyou calibration method to obtain camera calibration, and calculate the projection distance between the camera optical center and the world coordinate point according to the camera calibration; establish a projection relationship between the world coordinate and the pixel coordinate according to the projection distance and the camera calibration to obtain the three-dimensional world coordinate of the target center point; Step S5: Obtain consecutive frames of the second image, where the second image includes a first frame image and subsequent frame images; use the three-dimensional world coordinate of the first frame image as the initial coordinate, compare the three-dimensional world coordinates of the subsequent frame images with the initial coordinate to obtain a settlement value.

2. The non-contact monocular vision high-precision automatic settlement monitoring method according to claim 1, characterized in that, The process of data augmentation includes: Obtain the target image, where the target image includes targets at different distances, backgrounds, and angles; Add Gaussian noise and salt-and-pepper noise to the target image to generate a noisy image; Rotate the noisy image at different angles to generate multi-directional images; Perform brightness enhancement and brightness reduction operations on the multi-directional images to obtain the first data set.

3. The non-contact monocular vision high-precision automatic settlement monitoring method according to claim 1, characterized in that The preprocessing includes: Use the labelimg software to label the first data set in the yolo format to obtain a third data set; Divide the third data set according to expert experience to obtain a second data set; Among them, the second data set includes a training set, a validation set, and a test set. The training set is used to train the basic model; the validation set is used to evaluate the basic model after each round of training; the test set is used to test the YOLOv10s model.

4. A non-contact monocular vision high-precision automatic settlement monitoring method according to claim 1, characterized in that The YOLOv10s model includes: An input layer for resizing and normalizing the images in the second data set to obtain a third data set; A backbone network layer for extracting basic features of the third data set; A feature pyramid network layer for extracting high-level features according to the basic features and assigning weights to the high-level features by combining an adaptive feature fusion method to obtain a multi-scale feature map; An output layer for outputting the target position information and the confidence level of the third data set.

5. A non-contact monocular vision high-precision automatic settlement monitoring method according to claim 1, characterized in that, The process of obtaining the pixel coordinates includes: Perform grayscale conversion, image denoising, and binarization operations on the first image to obtain the second image; Use the Canny edge detection algorithm to extract the edge information of the second image and generate a binary edge map; Extract the edge point set of the target shape from the binary edge map, and use the adaptive threshold method to screen the edge point set to obtain the edge features; Use the least squares ellipse fitting algorithm and the adaptive weight adjustment method to perform ellipse fitting on the edge features to generate the elliptical shape of the target; Locate the center point of the target using the moment method based on the elliptical shape to obtain the pixel coordinates; 6. The non-contact monocular vision high-precision automatic settlement monitoring method according to claim 1, characterized in that The calculation process of the three-dimensional world coordinates includes: Obtain the camera calibration and calculate the projection distance. The camera calibration includes the internal parameter matrix, rotation matrix, and translation vector; Establish the projection relationship between the world coordinates and the pixel coordinates according to the camera imaging principle to obtain the projection function, which is expressed as: s*P C = M*(R*P w +T); where s is the projection distance; P C =(u, v, 1) is the pixel coordinate, where u and v are the normalized pixel coordinate values on the image plane; M is the intrinsic matrix; R is the rotation matrix; P w =(X, Y, Z) is the world coordinate, where X, Y, and Z are the coordinate values in the three-dimensional world respectively; T is the translation vector; Substitute the projection distance and the camera calibration into the projection function to obtain the three-dimensional world coordinates; 7. A non-contact monocular vision high-precision automatic settlement monitoring method according to claim 1, characterized in that, The calculation process of the three-dimensional world coordinates further includes: If there is more than one target in the target image, use the YOLOv10s model to identify each target one by one, and use geometric correction and depth estimation to correct the error of each three-dimensional world coordinate; 8. The non-contact monocular vision high-precision automatic settlement monitoring method according to claim 1, characterized in that, The calculation process of the projection distance includes: Set the vertical coordinate value of the world coordinates to 0; Calculate the projection distance according to the camera calibration and the pixel coordinates, which is expressed as: s = (Z + M2[2]) / M1[2]; where s is the projection distance; Z is the vertical coordinate value; M2[2] is the third component of the translation vector T projected into the camera coordinate system; and M1[2] is the pixel point P in the camera coordinate system C The converted depth information.

9. The non-contact monocular vision high-precision automatic settlement monitoring method according to claim 1, characterized in that The process of obtaining the settlement value includes: Obtain the first frame image and use the three-dimensional world coordinates of the first frame image as the initial coordinates; Obtain the subsequent frame images and use the three-dimensional world coordinates of the subsequent frame images as the subsequent coordinates; Subtract the subsequent coordinates in the vertical direction from the initial coordinates to obtain the settlement value; Correspond each frame timestamp with the settlement value to obtain settlement change data; Use the trained deep learning model to predict the future settlement value according to the settlement change data; Divide the future settlement value into geological uplift and geological subsidence, and divide the severity levels according to the severity of the geological uplift and the geological subsidence. The severity levels include the first level, the second level, and the third level; If the future settlement value is the second level, issue a warning; if the future settlement value is the third level, issue a warning and remind for manual intervention; 10. A non-contact monocular vision high-precision automatic settlement monitoring method according to claim 9, characterized in that, The process of obtaining the future settlement value includes: Use the sliding window method to input the settlement change data into the second basic model for training to obtain the trained deep learning model; Among them, if the length of the sliding window is w, the input data of the second basic model is w settlement change data; Input the latest settlement change data into the deep learning model for prediction to obtain the future settlement value, which is the settlement value at the next moment.

Citation Information

Cited By

  • Underground vehicle positioning method and system based on visual target information fusion

    CN120576776A

  • Structure settlement monitoring method and system based on image reconstruction

    CN120760675A

  • Image Reconstruction-Based Method and System for Monitoring Structure Settlement

    CN120760675B

  • Visual measurement method and system for structure settlement based on deep learning compensation

    CN121557957A