Civil construction period structure safety monitoring method and system based on image recognition

Abstract

The application provides a kind of civil construction period structure safety monitoring method and system based on image recognition, belong to computer vision and civil engineering safety monitoring combined technical field.The method comprises: S1 acquires binocular image data of construction site, constructs the data set of building material;S2. Based on the target detection and instance segmentation of building material in data set using improved Mask RCNN algorithm, obtain mask and spatial coordinates;S3. Stereoscopic matching is carried out to binocular image, and SGM algorithm is used to generate disparity map and convert into three-dimensional point cloud data;S4. Calculate the volume of building material based on point cloud data, and optimize the volume measurement result combined with the correction coefficient k of shooting angle θ;S5. Estimate the weight of building material according to material density and corrected volume, and output load distribution map.The method can replace traditional manual measurement, significantly improve the efficiency and safety of construction load statistics, and is suitable for real-time monitoring and early warning in complex construction site environment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision combined with civil engineering safety monitoring, and particularly relates to a civil construction period structure safety monitoring method and system based on image recognition, which is suitable for construction load real-time monitoring and structure safety early warning. BACKGROUND

[0002] During the construction stage of civil engineering, the concentrated stacking of building materials (such as steel bars, formworks, marbles, etc.) can easily lead to local load over-limit, causing structure collapse accidents. The traditional manual measurement method is low in efficiency and poor in accuracy, and cannot adapt to complex construction site environments. Although existing computer vision technology has been applied in the field of target detection, the following problems still exist:

[0003] There is a lack of special data sets and optimized algorithms for building materials, and the recognition accuracy is insufficient;

[0004] The binocular stereo matching algorithm has a large error in weak texture and occlusion scenes;

[0005] The perspective tilt caused by non-perpendicular shooting can significantly affect the volume measurement results, and the existing methods have not effectively solved the problem.

[0006] Therefore, there is an urgent need for a high-precision and automated building material weight monitoring scheme. SUMMARY

[0007] In view of the above problems, the application provides a civil construction period structure safety monitoring method and system based on image recognition, which combines an improved deep learning model and a building material weight estimation method based on binocular vision technology, to realize accurate statistics of construction load and safety early warning.

[0008] To solve the above technical problems, the technical scheme provided by the application is:

[0009] The application provides a civil construction period structure safety monitoring method and system based on image recognition, characterized by the following steps:

[0010] S1. Collecting binocular image data of the construction site to construct a data set of building materials;

[0011] S2. Target detection and instance segmentation of the building materials in the data set based on an improved Mask RCNN algorithm to obtain a mask and spatial coordinates;

[0012] S3. Stereo matching of the binocular images, generation of a disparity map by using an SGM algorithm, and conversion into three-dimensional point cloud data;

[0013] S4. Calculation of the volume of building materials based on the point cloud data, and optimization of the volume measurement results in combination with a correction coefficient k of the shooting angle θ;

[0014] S5. Estimate the weight of the building materials according to the material density and the corrected volume, and output the load distribution diagram.

[0015] Further, the improved Mask RCNN algorithm in step S2 includes:

[0016] S21 Network improvement: Insert an additional Identity module into the ResNet101 backbone network of Mask RCNN to increase the network depth and improve the feature extraction capability;

[0017] S22 Transfer training: Load the pre-training weights of the COCO dataset, freeze part of the bottom network parameters, and fine-tune the training on the building material dataset;

[0018] S23 Target detection: Input the binocular image into the trained model to output the prediction box, class confidence and mask area of the building material target;

[0019] S24 Spatial positioning: According to the center pixel coordinates of the prediction box, combined with the calibration parameters of the binocular camera, the three-dimensional spatial coordinates of the target are calculated by triangulation method.

[0020] Further, the specific implementation steps of the SGM algorithm in step S3 include:

[0021] Cost calculation: The pixel matching cost C(p,d) is calculated using Census transformation, and the formula is:

[0022] C(p,d) = ∑ q∈Np XOR(Census(q), Census(q d ))

[0023] Where N p is the neighborhood window of pixel p, and q d is the corresponding pixel of the transformed pixel q under the disparity d;

[0024] Cost aggregation: The cost value is accumulated along 16 paths in horizontal, vertical and diagonal directions;

[0025] Disparity calculation: The disparity corresponding to the minimum cost value is selected by the winner-takes-all (WTA) strategy;

[0026] Disparity optimization: Mis-matching points are removed by left-right consistency check, and holes are filled by median filtering.

[0027] Further, the specific method of disparity optimization includes:

[0028] Cross-validation is performed on the disparity maps before and after optimization, and pixels that satisfy ∣d left -d right ∣≤1 are retained;

[0029] The median filter window of 5x5 pixels is used to fill the hole region;

[0030] The disparity accuracy is improved by sub-pixel interpolation, and the interpolation formula is:

[0031]

[0032] Wherein, d sub : sub-pixel optimized disparity value, d: candidate disparity value; C(d), C(d-1), C(d+1) represent the generation value corresponding to the disparity value d, d-1 and d+1; d left : pre-optimized disparity value, d right : post-optimized disparity value.

[0033] Further, the method for determining the shooting angle correction coefficient in step S4 comprises:

[0034] The standard cube is shot by the binocular camera at different shooting angles θ to obtain the corresponding measured volume V_measured;

[0035] The volume error ΔV=|V_measured-V_standard| under each angle is calculated, wherein V_standard is the reference volume under the vertical angle, i.e. θ=90°.

[0036] The linear relationship k=αθ+β between the angle θ and the correction coefficient k is fitted by the least square method, wherein α and β are experimental calibration parameters.

[0037] Further, the specific form of the linear relationship is:

[0038] k=0.0085θ+0.235

[0039] Wherein θ is the shooting angle, i.e. the angle between the camera optical center line and the horizontal plane, and the value range of θ is 20°≤θ≤85°.

[0040] Further, the specific method for volume calculation in step S4 is:

[0041] The three-dimensional point cloud is projected onto a two-dimensional plane, the grid elements are divided, and the average height h avg， i;

[0042] The volume of all elements is accumulated, and the formula is:

[0043] V measured =∑ N i＝1 (A cell ×h avg,i )

[0044] Wherein Acell is the area of a grid cell, and N is the total number of grids.

[0045] Further, the step S4 further comprises error compensation for the irregular building materials.

[0046] When it is detected that the point cloud data has occlusion or missing, the data is supplemented through multi-view image acquisition;

[0047] The multi-view point cloud is fused, and the volume is recalculated by using the numerical integration method.

[0048] Further, the step S5 further comprises:

[0049] The weight estimation result is mapped to the finite element model, and the stress and deformation of the component are calculated.

[0050] When the stress exceeds 90% of the designed resistance, a pre-warning is triggered, and when the stress exceeds 100% of the designed resistance, an alarm is triggered.

[0051] In another aspect, the present application also claims to protect a building material weight estimation system based on any of the foregoing methods, comprising:

[0052] The binocular image acquisition module is used for fixed-angle shooting and transmitting construction site images;

[0053] The building material recognition module outputs target masks and coordinates based on the improved Mask RCNN algorithm.

[0054] The stereo matching module generates a disparity map and reconstructs a three-dimensional point cloud by using the SGM algorithm.

[0055] The error correction module optimizes the measurement results by using the angle-volume relationship.

[0056] The load analysis module calculates the weight and generates a load distribution map in combination with the density.

[0057] In another aspect, the present application also claims to protect an electronic device comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the program to realize the steps of any of the foregoing methods.

[0058] In another aspect, the present application also claims to protect a computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to realize the steps of any of the foregoing methods.

[0059] Compared with the prior art, the civil construction period structure safety monitoring method and system based on image recognition have the following beneficial technical effects:

[0060] Improved recognition accuracy: the improved Mask RCNN has an average confidence of 87.58% on the test set, which is 12% higher than the original model.

[0061] Volume measurement is efficient: the SGM algorithm has a low void rate of less than 5% in building material images, and the calculation speed is 1.78 seconds / frame;

[0062] Error is significantly reduced: after angle correction, the volume error is compressed from 15.3% to within 5%;

[0063] Realize automatic monitoring, replace manual inspection, and reduce the risk of safety accidents. BRIEF DESCRIPTION OF DRAWINGS

[0064] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings described below are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0065] Figure 1 The flowchart of the civil construction period structure safety monitoring method and system based on image recognition provided by the embodiment of the present application.

[0066] Figure 2 The data set pictures provided by the embodiment of the present application, wherein (a) is a typical steel bar data set sample; (b) is a typical marble data set sample; (c) is a typical support template data set sample.

[0067] Figure 3 The building material labels and mask pictures provided by the embodiment of the present application, wherein (a) is a Labelme annotation example; (b) is a marble label and mask picture; (c) is a support template label and mask picture; (d) is a steel bar label and mask picture.

[0068] Figure 4 The process of building material recognition provided by the embodiment of the present application;

[0069] Figure 5 The structural module diagram of the computer vision-based building material identification and weight estimation system electronic device provided by the present application. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0071] The application will be further described below in conjunction with examples:

[0072] Implementation device

[0073] Binocular camera: model HBV-1780S2.0, baseline distance 60mm, focal length 2.1mm, field of view angle 100°, resolution 1280x720;

[0074] Calibration tool: 10x7 black and white chessboard calibration board, each square size 28mmx28mm;

[0075] Labeling software: Labelme 4.5.7, supporting polygon mask labeling;

[0076] Computing platform: cloud server equipped with NVIDIA Tesla T4 GPU, memory 56GB, hard disk 80GB.

[0077] Implementation step S1 construction site binocular image acquisition and building material dataset construction method

[0078] 1. Binocular camera installation and calibration

[0079] The various data of the application are respectively from Jiangxi Lushan high-speed station expansion project, Beijing urban construction Langfang project and China Railway Real Estate Beijing project. Deploy a binocular camera with a baseline distance of 60mm (model HBV-1780S2.0), camera focal length 2.1mm, field of view angle 100°, resolution 1280x720. In order to ensure the measurement accuracy, first calibrate the binocular camera: use a 10x7 chessboard calibration board to take multiple images at different angles, and calculate the intrinsic matrix K, extrinsic matrix [R|t] and distortion coefficient through the Matlab binocular calibration toolbox. The calibration results show that the average reprojection error is stable within 0.15 pixels (Table 1), which meets the engineering accuracy requirement.

[0080] Table 1 Calibration results of each group of images

[0081]

[0082] The unit of focal length and reprojection error is pixel. Since the focal length difference and reprojection error are both small, the calibration results of the 4th group of images are selected as the final results of the binocular camera parameters for subsequent measurement work. The calibration parameter results of the 4th group of images are shown in Table 2:

[0083] Table 2 Camera parameters used in this paper

[0084]

[0085] From the table data can be seen, the rotation matrix of binocular camera is close to the value of unit matrix, the first component of translation matrix is-60.6999mm, which is close to the baseline length of the camera. The above data are close to the actual parameter values of the binocular camera used in this paper, and basically meet the use requirements of the measurement work in the following.

[0086] 2. Construction site image acquisition

[0087] Scene selection: Mask RCNN training dataset includes steel bars, marbles and support template pictures. As shown in Figure 2 , some typical training set samples are shown. Steel bars are collected in bundles, templates are stacked in layers, and marble plates are laid out in scenes;

[0088] Shooting requirements:

[0089] At least 300 images of each material are taken, covering different light (sunny noon, cloudy day, night light supplement), shielding (instrument shielding, worker operation) and stacking form (single layer paving, multi-layer interlacing);

[0090] The target area accounts for more than 30% of the image area, and the background includes typical construction elements such as scaffolding and concrete structure;

[0091] Data storage: images are saved in PNG format, and the naming rule is "category_scene number_left-right identifier" (such as "rebar_001_left.png").

[0092] 3. Data annotation and enhancement

[0093] Mask annotation: use Labelme tool to draw polygon vertices along the building material contour, generate JSON format annotation file, and the class label is "rebar" "moban" "marble", see Figure 3 ;

[0094] Data enhancement: randomly apply rotation (±15°), horizontal flip, brightness adjustment (±20%) and contrast change (0.8-1.2 times) to the original image, expand the dataset to 3 times the original size;

[0095] Dataset division: randomly divided into training set (steel bar 360, template 288, marble 315) and test set (steel bar 40, template 32, marble 35) according to the ratio of 9:1.

[0096] Step S2: Improve the building material recognition and positioning network structure of Mask RCNN

[0097] The labeled data set is input into the Mask RCNN convolutional neural network for model training. When the data set and related files are made, the computer system is Windows, and because the computer hardware cannot support the training of the model, cloud computing power is selected for training. The cloud computing power computer environment for building the Mask RCNN convolutional neural network in this application is python3.8, cuda11.2 and cudnn9. The IDE is Pycharm2022. The computer hardware configuration is: the processor is Intel Xeon Processor, the GPU is Tesla T4, the memory is 56G, and the hard disk is 80G.

[0098] The algorithm flow of building material target detection by the Mask RCNN network is as shown in Figure 4

[0099] In the improved Mask RCNN model, the ResNet101 backbone network is inserted into an additional Identity module to enhance the feature extraction capability. Each Identity module contains two 3x3 convolution layers, followed by batch normalization (BatchNorm) and ReLU activation function, and the input and output sizes are kept consistent through a skip connection. The model loads the COCO data set pre-training weight, freezes the first 50 layer parameters, and uses the Adam optimizer (initial learning rate 1e-4) to fine-tune the building material data set. During the training process, the loss function integrates the classification loss (Focal Loss), the regression loss (Smooth L1) and the mask loss (cross entropy), and finally achieves an average confidence of 81.25% for steel bars, 86.71% for templates and 94.78% for marbles on the test set (Table 3).

[0100] Table 3 Average confidence of building material prediction results

[0101]

[0102] Table 3-1 is the average confidence of the prediction results of the three types of building materials. The results show that when the objects in the image are in a relatively regular environment, the prediction confidence result is relatively higher.

[0103] In the inference stage, the corrected left view is input into the model, and the prediction box, mask area and center pixel coordinates (u l , v l ) of the target are output. Combined with the double target determination parameters, the three-dimensional coordinates (X, Y, Z) are calculated, and the formula is:

[0104]

[0105] Where f = 2.1 mm, B = 60 mm, c x = 640, c y ​= 360, d = u l - u r The average accuracy of the three-dimensional coordinates of 10 groups of targets measured was 90.18% (Table 4).

[0106] Table 4 Space coordinate measurement results and accuracy

[0107]

[0108] The first five groups in the table are data measured by object 1, and the last five groups are data measured by object 2. The pixel coordinate values of the prediction box and the center point of the two objects are obtained by the Mask RCNN algorithm, the disparity corresponding to the center point is calculated by the stereo matching algorithm, and finally the pixel coordinates of the center point of the target object in the image are converted into three-dimensional coordinate values in the coordinate system with the optical center of the left eye camera as the origin by the aforementioned coordinate conversion method. The related principles of the stereo matching algorithm will be described in detail later.

[0109] Step S3: SGM stereo matching and three-dimensional point cloud generation

[0110] 1. Stereo matching: Stereo matching is the core link of volume calculation. In this embodiment, the SGM (semi-global matching) algorithm is used, and the specific process is as follows:

[0111] Cost calculation: For each pixel p in the pre-conversion image and the post-conversion image, a 9x9 neighborhood window is extracted, and a binary bit string is generated by Census transformation. If the gray value of the neighborhood pixel q is greater than that of the center pixel, it is marked as 1, otherwise as 0. The Hamming distance of the corresponding pixels of the pre-conversion image and the post-conversion image is calculated as the matching cost C(p, d), and the formula is:

[0112] C(p, d) = ∑ q∈Np XOR(Census(q), Census(q d ))

[0113] Where q d is the corresponding pixel in the post-conversion image under the disparity d.

[0114] Cost aggregation: The path aggregation formula is introduced along the 16-direction path to accumulate the cost value, and the smoothing penalty coefficients P1 = 50 (disparity change 1) and P2 = 150 (disparity change more than 1) are introduced.

[0115] Disparity optimization: The winner-takes-all (WTA) strategy is used to select the disparity corresponding to the minimum cost value to generate the initial disparity map. The front and back consistency check is performed on the optimized disparity map, and the pixels that satisfy |d left -d right |≤1 are retained;

[0116] A 5x5 pixel median filter window is used to fill the hole region;

[0117] And the disparity accuracy is improved by sub-pixel interpolation, the interpolation formula is:

[0118]

[0119] Where d sub : the disparity value after sub-pixel optimization, d: the candidate disparity value; C(d), C(d-1), C(d+1) represent the generation value corresponding to the disparity value d, d-1 and d+1; d left : the disparity value before optimization, d right : the disparity value after optimization.

[0120] Point cloud generation:

[0121] Coordinate conversion: convert the disparity map d sub into a three-dimensional point cloud (PLY format), and remove outliers by statistical filtering (mean K=50, standard deviation multiplier 1.0).

[0122] Step S4: volume calculation and angle correction

[0123] Volume calculation:

[0124] Projection integral method: project the point cloud onto a 2mm x 2mm grid, and calculate the average height h avg，i of each unit.

[0125] Accumulate the volume of all units, the formula is:

[0126] V measured =∑ N i＝1 (A cell ×h avg,i )

[0127] Where A cell is the area of the grid unit, and N is the total number of grids.

[0128] Angle correction: in the non-perpendicular shooting scene, the volume underestimation problem caused by the inclination angle θ of the camera optical center line needs to be corrected, and the corrected volume is:

[0129] V corrected = k × V measured

[0130] Correction coefficient k calculation: shoot a standard cube with different shooting angles θ by a binocular camera, and obtain the corresponding measured volume V_measured;

[0131] Calculate the volume error ΔV=|V_measured-V_standard| under each angle, where V_standard is the reference volume under the vertical angle, i.e. θ=90°.

[0132] The linear relationship between the angle θ and the correction coefficient k is fitted by the least square method, where α and β are experimental calibration parameters.

[0133] Further, the linear relationship is specifically:

[0134] k = 0.0085θ + 0.235

[0135] Wherein θ is the shooting angle, i.e. the angle between the camera optical center line and the horizontal plane, and the value range of θ is 20°≤θ≤85°.

[0136] In another embodiment of the present application, referring to Figure 5 The present application also provides an electronic device, in particular:

[0137] The electronic device can include a processor with one or more processing cores, a memory with one or more computer readable storage media, a power supply, an input unit, and the like. Among them:

[0138] The processor is the control center of the electronic device, which connects all parts of the electronic device through various interfaces and lines, executes the software programs and / or modules stored in the memory, and calls the data stored in the memory, to perform various functions and process data of the electronic device, thereby overall monitoring the electronic device. Optionally, the processor can include one or more processing cores; the processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor and the like, preferably, the processor can integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface and application programs, etc., and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor can also not be integrated into the processor.

[0139] The memory can be used to store software programs and modules, and the processor executes various function applications and data processing by running the software programs and modules stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, application programs required by at least one function (such as a sound playing function, an image playing function, etc.), and the like; and the data storage area can store data created according to the use of the electronic device, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device. Accordingly, the memory can also include a memory controller to provide the processor with access to the memory.

[0140] The electronic device also includes a power supply for powering the various components. Preferably, the power supply can be connected to the processor logic through a power management system, so that the power management system can be used to manage charging, discharging, and power consumption management, etc. The power supply can also include one or more direct current or alternating current power sources, a recharging system, a power failure detection circuit, a power converter or inverter, a power status indicator, and the like.

[0141] The electronic device can also include an input unit, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function controls.

[0142] Although not shown, the electronic device can also include a display unit, etc., which will not be described here. In particular, in the embodiment, the processor in the electronic device loads the executable file corresponding to the process of one or more application programs into the memory according to the following instructions, and runs the application programs stored in the memory by the processor, thereby realizing various functions.

[0143] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructions, or by instructions controlling related hardware, which can be stored in a computer readable storage medium and loaded and executed by a processor.

[0144] In some embodiments of the present application, the present application also provides a computer readable storage medium, which can include a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc. A computer program is stored thereon, and the computer program is loaded by a processor to execute the steps in the order processing method provided by the embodiments of the present application.

[0145] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit the same; and although the present application has been described in detail with reference to the foregoing embodiments, it should be appreciated by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features thereof can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An image recognition-based monitoring method for structural safety during civil construction, characterized by, The method comprises the following steps: S1. Collecting binocular image data of the construction site and constructing a dataset of building materials; S2. Target detection and instance segmentation of building materials in the dataset based on an improved Mask RCNN algorithm to obtain a mask and spatial coordinates; S3. Stereoscopic matching of the binocular image, generation of a disparity map by using an SGM algorithm, and conversion into three-dimensional point cloud data; the specific implementation steps of the SGM algorithm in the step S3 comprise: Cost calculation: calculation of pixel matching cost C(p, d) by using Census transformation, with the formula being: C(p,d) =∑ q∈Np XOR(Census(q), Census(q d )) where N p is a neighborhood window of pixel p, q d is the corresponding pixel of transformed pixel q at disparity d; Cost aggregation: accumulation of cost values along 16 paths in horizontal, vertical and diagonal directions; Disparity calculation: selection of the disparity corresponding to the minimum cost value by using a winner-takes-all (WTA) strategy; Disparity optimization: elimination of mis-matching points by using left-right consistency checking and filling of holes by using median filtering; S4. Calculation of the volume of building materials based on the point cloud data and optimization of the volume measurement result in combination with a correction coefficient k of the shooting angle θ; S5. Estimation of the weight of building materials according to the material density and the corrected volume and output of a load distribution map.

2. The method of claim 1, wherein, The improved Mask RCNN algorithm in the step S2 comprises: S21 Network improvement: insertion of an additional Identity module into the ResNet101 backbone network of Mask RCNN to increase the network depth and improve the feature extraction capability; S22 Migration training: loading of pre-training weights of a COCO dataset, freezing of part of the bottom network parameters, and fine-tuning training of the building material dataset; S23 Target detection: input of the binocular image into the trained model to output the prediction box, class confidence and mask area of the building material target; S24 Spatial positioning: calculation of the three-dimensional spatial coordinates of the target by using the triangular measurement method according to the center pixel coordinates of the prediction box and in combination with the calibration parameters of the binocular camera.

3. The method of claim 1, wherein, The specific method of disparity optimization comprises: The disparity map before and after optimization is cross-validated, and the pixels satisfying |d left -d right |≤1 are retained; Filling of holes in the empty area by using a 5×5 pixel median filtering window; Improvement of the disparity accuracy by using sub-pixel interpolation, with the interpolation formula being: Wherein, d sub : disparity value after sub-pixel level optimization, d: candidate disparity value; C(d), C(d-1), C(d+1) represent the generation value corresponding to disparity value d, d-1 and d+1; d left disparity value before optimization, d right disparity value after optimization.

4. The method of claim 1, wherein, The determination method of the correction coefficient k in the step S4 comprises: Shooting of a standard cube at different shooting angles θ by using the binocular camera to obtain the corresponding measured volume V_measured; Calculation of the volume error ΔV = |V_measured-V_standard| at each angle, wherein V_standard is the reference volume at the vertical angle, i.e. θ = 90°; Fitting of the linear relationship between the angle θ and the correction coefficient k by using the least square method, with the formula being k = αθ + β, wherein α and β are experimental calibration parameters.

5. The method of claim 4, wherein, The specific form of the linear relationship is: k = 0.0085θ + 0.235 Wherein θ is the shooting angle, i.e. the included angle between the camera optical center line and the horizontal plane, and the value range of θ is 20° ≤ θ ≤ 85°.

6. The method of claim 1, wherein, The specific method of volume calculation in the step S4 is: Projecting a three-dimensional point cloud to a two-dimensional plane, dividing grid cells and calculating the average height h of each cell avg，i ; Accumulation of all unit volumes, with the formula being: V measured =∑ N i=1 (A cell ×h avg,i ) where A cell is the area of a grid cell, and N is the total number of grids.

7. The method of claim 1, wherein, The step S4 further comprises error compensation for irregular building materials: When it is detected that the point cloud data has occlusion or loss, the data is supplemented by multi-view image acquisition; Fusion of multi-view point clouds and recalculation of the volume by using the numerical integration method.

8. The method of claim 1, wherein, The step S5 further comprises: mapping the weight estimation results to the finite element model to calculate the stress and deformation of the component; triggering a warning when the stress exceeds 90% of the design resistance, and triggering an alarm when it exceeds 100%.

9. A building material weight estimation system based on the method of any one of claims 1-8, characterized by, Comprise: a binocular image acquisition module for fixed-angle shooting and transmission of construction site images; a building material recognition module that outputs target masks and coordinates based on an improved Mask RCNN algorithm; a stereo matching module that generates a disparity map and reconstructs a three-dimensional point cloud using the SGM algorithm; an error correction module that optimizes the measurement results through an angle-volume relationship; a load analysis module that calculates the weight and generates a load distribution map in combination with the density.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program, when executed by the processor, is arranged to perform the method of any one of claims 1 to 9. The processor executes the program to implement the steps of the method of any one of claims 1-8.

11. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the method of any one of claims 1-8.

Images