Building main body structure construction method optimization method and system based on AI visual recognition

Abstract

This invention relates to the field of machine vision technology, specifically to a method and system for optimizing the construction method of building main structures based on AI visual recognition. The method includes generating a time-varying probability grid map, assessing pump truck path risks based on the grid map and generating an incremental sequence of liquid surface vertical elevation, processing the sequence using a convolutional temporal structure to generate pumping adjustment signals, collecting point clouds to construct a two-dimensional deviation distribution matrix, and generating construction records based on frequency domain transformation and secondary keel spacing matching. In this invention, by constructing a dynamic diffusion grid space to map the movement of suspended objects into a time-varying probability map, dynamic collision risk quantification is achieved. Depth perception is used to monitor liquid surface elevation, and a convolutional model is used to fit lateral pressure dissipation characteristics to achieve adaptive adjustment of pumping frequency. Frequency domain transformation is used to analyze the dominant frequency of the deviation matrix and match the keel spacing, accurately identifying stiffness-failed components and optimizing support spans, eliminating blind spots in manual monitoring and improving construction accuracy and management efficiency.

Description

Technical Field

[0001] This invention relates to the field of machine vision technology, and in particular to a method and system for optimizing the construction methods of building main structures based on AI visual recognition. Background Technology

[0002] The field of machine vision technology involves a technical system that uses cameras and computers to simulate the human visual system to perceive, recognize, and understand three-dimensional scenes in the objective world. Its core encompasses processing flows such as image acquisition, preprocessing, feature extraction and matching, 3D reconstruction, target detection, and tracking. It acquires target images through optical devices and converts them into digital signals, then uses image processing systems for computation and analysis to achieve functions such as target size measurement, positioning guidance, defect detection, and behavior analysis. Among these methods, the traditional AI-based visual recognition-based construction method optimization for building structures involves installing fixed-point video surveillance cameras at tower cranes or key work nodes on the construction site to collect real-time video images of the construction site. The video stream is then aggregated to a display screen in the monitoring center via transmission cables or wireless networks. Management personnel monitor the construction progress and on-site safety by visually observing the screen images. During construction method acceptance or correction, manual labor is used to carry paper drawings and measuring tools such as tape measures to conduct on-site measurements and records. Based on manual experience, the differences between the actual progress and the planned progress are compared to formulate subsequent construction adjustment plans.

[0003] Current technology relies on fixed-point surveillance cameras to collect images and aggregate them to a monitoring center. Managers need to constantly monitor the construction progress and safety by visually observing the screen. This manual monitoring mode is difficult to detect subtle dynamic risks and is prone to causing visual fatigue. In the process of construction method acceptance and correction, workers need to carry paper drawings and physical measuring tools to conduct on-site measurements and rely on subjective experience to compare progress differences. This not only results in low efficiency in obtaining on-site measurement data but also fails to establish a full-cycle digital mapping, leading to delays in the formulation of construction adjustment plans and a lack of accurate quantitative data support. Summary of the Invention

[0004] To achieve the above objectives, the present invention adopts the following technical solution: an optimization method for construction techniques of building main structures based on AI visual recognition, comprising the following steps:

[0005] S1: Acquire video frame sequence of the object being lifted by the tower crane, calculate pixel displacement to generate the centroid displacement vector of the object being lifted, construct a grid space based on the velocity direction, calculate the diffusion radius and probability based on the speed and direction of the object being lifted and map them to generate a time-varying probability grid map.

[0006] S2: Obtain the planned path of the concrete pump truck boom, map it to the time-varying probability grid map to perform a collision risk assessment, respond to the assessment results and drive the camera to acquire a depth map, segment the liquid surface area and calculate the height difference, and aggregate to generate an incremental sequence of vertical elevation of the liquid surface.

[0007] S3: Input the vertical elevation increment sequence of the liquid surface into the convolutional time series processing structure, use the exponential decay weight to fit the side pressure dissipation characteristics, obtain the equivalent side pressure index by weighted summation of historical data, compare the equivalent side pressure index with the bearing standard, and generate the pumping stroke frequency adjustment signal.

[0008] S4: In response to the termination of the transmission of the pumping stroke frequency adjustment signal, collect three-dimensional point cloud data of the wall surface, fit and calculate the deviation angle between the normal vector and the design normal vector, map it to two-dimensional plane coordinates, and construct a two-dimensional deviation distribution matrix.

[0009] S5: Extract the dominant frequency component from the frequency domain transformation of the two-dimensional deviation distribution matrix, convert it into physical wavelength and compare it with the spacing of the secondary keel, identify the failed components, calculate the maximum allowable spacing configuration data of the support components based on the elastic modulus, and generate structural construction record data.

[0010] As a further embodiment of the present invention, the time-varying probability grid map includes a grid coordinate index, a spatial occupancy probability value, and a dynamic diffusion boundary marker; the liquid surface vertical elevation increment sequence includes a sampling timestamp, a liquid surface vertical height difference, and a pouring area marker; the pumping stroke frequency adjustment signal includes a stroke frequency target value, a pressure safety margin coefficient, and a pumping rate control command code; the two-dimensional deviation distribution matrix includes a plane mapping coordinate, a normal deviation angle value, and a local flatness level marker; and the structural construction record data includes a stiffness failure component positioning index, optimized secondary keel spacing parameters, and frequency domain feature verification results.

[0011] As a further aspect of the present invention, the steps for obtaining the time-varying probability grid map are specifically as follows:

[0012] S101: Acquire the video frame sequence of the tower crane hoisting object recorded by the visual acquisition device, call the preset camera intrinsic parameter matrix and spatial calibration coefficient, detect and lock the outline region of the hoisting object in the video frame, extract the pixel coordinates of the geometric center of the outline region, calculate the pixel coordinate difference of the geometric center between the current frame and the previous frame, and combine the frame rate parameter of the acquisition device to convert the pixel coordinate difference into a velocity vector to generate the displacement vector of the hoisting object's center of gravity.

[0013] S102: Call the displacement vector of the suspended object's center of gravity, analyze the vector magnitude as the instantaneous motion rate of the suspended object, analyze the vector angle as the motion direction angle, construct a two-dimensional grid coordinate system with the current position of the suspended object as the origin, calculate the risk diffusion distance based on the instantaneous motion rate, determine the angular range of the diffusion fan-shaped area in combination with the motion direction angle, use the risk diffusion distance as the radial length limit of the fan-shaped area, and generate the radius of the grid diffusion area.

[0014] S103: For each grid cell within the radius coverage of the grid diffusion area, obtain the grid center point coordinates, calculate the Euclidean distance between the grid center point coordinates and the predicted position of the suspended object, and map it as a collision risk weight. Normalize the collision risk weight to obtain the occupancy probability value, establish the correspondence between the occupancy probability value and the grid coordinate index, and generate a time-varying probability grid map.

[0015] As a further aspect of the present invention, the process of calling the preset camera intrinsic parameter matrix and spatial calibration coefficients specifically includes:

[0016] Collect a multi-angle static image sequence containing a checkerboard calibration board, calculate the spatial gradient magnitude of the image pixel grayscale in the horizontal and vertical directions, select the pixel positions where the gradient magnitude simultaneously presents a local maximum in the orthogonal direction as calibration corner points, and perform sub-pixel level interpolation positioning.

[0017] A geometric mapping equation describing the projection relationship from three-dimensional spatial coordinates to two-dimensional image plane coordinates is established. The actual coordinates of the calibrated corner points are substituted into the equation to construct a reprojection residual function that characterizes the deviation between the calculated projection point and the actual corner point position. The focal length, principal point coordinates, and radial distortion coefficient are updated through multiple iterations to minimize the value of the reprojection residual function and construct the camera intrinsic parameter matrix.

[0018] Identify a rectangular reference object with a known physical side length within the working plane, extract the vertex coordinates of the rectangular reference object on the image plane, perform distortion correction on the vertex coordinates based on the radial distortion coefficient in the camera intrinsic parameter matrix, calculate the pixel Euclidean distance between adjacent vertices after correction, calculate the ratio of the known physical side length to the pixel Euclidean distance, establish it as the conversion benchmark between image pixel units and physical space length, and generate the spatial calibration coefficient.

[0019] As a further aspect of the present invention, the step of obtaining the vertical elevation increment sequence of the liquid surface specifically includes:

[0020] S201: Obtain the set of coordinates for the planned path of the concrete pump truck boom, project and map the three-dimensional coordinate points in the set of planned path coordinates onto the two-dimensional index of the time-varying probability grid map, retrieve the space occupancy probability value in the corresponding grid, conduct a collision risk assessment, and output a safe operation permit instruction based on the assessment results.

[0021] S202: In response to the safety operation permit instruction, start the depth camera, acquire the surface depth image of the liquid concrete being poured, calculate the spatial gradient amplitude of each pixel in the depth image through edge detection, filter the pixels with spatial gradient amplitude below the flatness threshold and aggregate them into connected components, extract the depth measurement values ​​of all pixels in the connected components, establish the association mapping between the depth measurement values ​​and the acquisition time, and generate a time-segmented liquid surface depth data set.

[0022] S203: Call the time-segmented liquid surface depth data set, calculate the arithmetic mean of the pixel depth measurement values ​​in the connected domain at each acquisition time as the instantaneous liquid surface height at the target time, select two adjacent instantaneous liquid surface heights in the time series to perform numerical difference calculation, obtain the vertical height change within a unit time step, reorganize the vertical height change according to the time order, and generate a liquid surface vertical elevation increment sequence.

[0023] As a further aspect of the present invention, the process of filtering pixels whose spatial gradient magnitude is lower than the flatness threshold specifically includes:

[0024] In the initialization phase before the monitoring task is started, a stationary horizontal plane in the construction site is selected as the reference object. The reference depth image of the reference object is collected, and the spatial gradient amplitude of each pixel in the reference depth image is calculated point by point. Statistical analysis is performed on all spatial gradient amplitude samples to construct a Gaussian normal distribution model. The mathematical expectation and standard deviation parameters of the model are extracted, and the sum of the mathematical expectation and three times the standard deviation is calculated and set as the flatness threshold that covers the inherent measurement noise range of the sensor.

[0025] During real-time monitoring, the pixel spatial gradient amplitude of the depth image of the surface of the liquid concrete being poured is calculated. The amplitude is compared with the flatness threshold. Pixels with spatial gradient amplitudes lower than the flatness threshold are selected, and the pixels are determined to belong to the flat fluid surface area. They are marked as valid liquid surface data points and the corresponding depth measurement values ​​are retained.

[0026] As a further aspect of the present invention, the step of obtaining the pumping stroke frequency adjustment signal specifically includes:

[0027] S301: Obtain the initial setting time parameter and hardening rate coefficient of concrete material, construct a numerical decay function that decreases exponentially with time based on the hardening rate coefficient, calculate the signal retention weight value corresponding to each historical moment based on the sampling time step, construct a time-series processing structure containing multi-layer causal convolution operation units, map the signal retention weight value to the weight parameter matrix of the convolution kernel, and generate a time-varying decay convolution kernel for lateral pressure.

[0028] S302: Call the liquid surface vertical elevation increment sequence into the time series processing structure, use the side pressure time-varying decay convolution kernel to perform sliding window convolution operation on the sequence data, multiply each historical increment data point covered by the window by the corresponding weight parameter, perform cumulative calculation on all weighted data points, analyze the side pressure superposition effect, output the pressure estimate value at the current moment, and generate the equivalent side pressure index.

[0029] S303: Based on the equivalent lateral pressure index, retrieve and compare the preset standard values ​​of the ultimate bending capacity of the support member, calculate the numerical deviation as a safety margin index, adjust the target operating frequency of the pumping stroke, convert the target operating frequency into equipment control commands, and generate a pumping stroke frequency adjustment signal.

[0030] As a further aspect of the present invention, the step of obtaining the two-dimensional deviation distribution matrix specifically includes:

[0031] S401: Real-time transmission status monitoring is performed on the pumping stroke frequency adjustment signal. When the pumping operation ends, the laser scanner is triggered to perform a panoramic ranging scan on the surface of the completed wall, collect laser echo data and convert it into three-dimensional spatial coordinates, remove outlier noise, establish an index relationship between the three-dimensional coordinate points and the scanning intensity, and generate a point cloud set of the wall surface to be measured.

[0032] S402: For the three-dimensional coordinate points in the point cloud set of the wall surface to be measured, construct a spherical neighborhood with the coordinate points as the center and retrieve neighboring points. Construct the covariance matrix of the local plane based on the least squares method, analyze the eigenvalues ​​and eigenvectors of the matrix, select the eigenvector corresponding to the smallest eigenvalue as the normal direction, perform vector magnitude normalization operation, establish the mapping relationship between coordinate points and normal vectors, and generate the measured local unit normal vector set;

[0033] S403: Call the measured local unit normal vector set, calculate the cosine value of the angle between each measured normal vector and the ideal design plane normal vector, obtain the angle deviation value through inverse cosine operation, extract the horizontal and vertical components of the three-dimensional coordinate points on the wall facade, map them into the row and column numbers of the two-dimensional grid, fill the angle deviation value into the grid, and generate a two-dimensional deviation distribution matrix.

[0034] As a further aspect of the present invention, the steps for obtaining the structural construction record data are specifically as follows:

[0035] S501: Call the two-dimensional deviation distribution matrix to perform a feature transformation from the spatial domain to the frequency domain, calculate the modulus square of the complex spectrum, construct the power spectral density matrix, lock the dominant frequency component in the energy concentration, calculate the reciprocal of the dominant frequency component based on the spatial domain sampling frequency, obtain the corresponding spatial period length, and generate the dominant deformation physical wavelength.

[0036] S502: Obtain the secondary keel design spacing parameters in the architectural design drawings, perform numerical matching verification between the dominant deformation physical wavelength and the spacing parameters, analyze the correlation between the wall deformation cycle and the skeleton arrangement, determine the type of support component that affects the deformation, lock the target support component with stiffness failure, and generate a stiffness failure component location index.

[0037] S503: For the target component of the stiffness failure component positioning index, obtain the elastic modulus constant of the corresponding material and the target flatness requirement, calculate the maximum allowable spacing configuration data of the support components, integrate the stiffness failure analysis results and the spacing configuration data, and generate structural construction record data.

[0038] A construction method optimization system for building main structures based on AI visual recognition includes:

[0039] The dynamic risk raster mapping module acquires video frame sequences of objects lifted by tower cranes, calculates pixel displacement to generate the centroid displacement vector of the object, constructs a raster space based on the velocity direction, calculates and maps the diffusion radius and probability based on the speed and direction of the object's movement, and generates a time-varying probability raster map.

[0040] The liquid surface increment sequence extraction module obtains the planned path of the concrete pump truck arm, maps it to the time-varying probability grid map for collision risk assessment, responds to the assessment result and drives the camera to acquire depth map, segments the liquid surface area and calculates the height difference, and aggregates to generate a liquid surface vertical elevation increment sequence.

[0041] The pumping frequency feedback control module inputs the vertical elevation increment sequence of the liquid surface into the convolutional time series processing structure, uses exponential decay weights to fit the side pressure dissipation characteristics, obtains the equivalent side pressure index by weighted summation of historical data, compares the equivalent side pressure index with the bearing standard, and generates a pumping stroke frequency adjustment signal.

[0042] The two-dimensional deviation matrix construction module, in response to the termination state of the transmission of the pumping stroke frequency adjustment signal, collects three-dimensional point cloud data of the wall surface, fits and calculates the deviation angle between the normal vector and the design normal vector, maps it to two-dimensional plane coordinates, and constructs a two-dimensional deviation distribution matrix.

[0043] The structural stiffness diagnosis and recording module extracts the dominant frequency components from the frequency domain transformation of the two-dimensional deviation distribution matrix, converts them into physical wavelengths, compares them with the spacing of the secondary keel, identifies failed components, calculates the maximum allowable spacing configuration data of the support components based on the elastic modulus, and generates structural construction record data.

[0044] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0045] In this invention, the centroid displacement vector is calculated by acquiring tower crane video frames and a dynamic diffusion grid space is constructed. The motion of the suspended object is mapped into a time-varying probability grid map, realizing a dynamic collision risk quantification assessment. Depth perception is used to monitor the changes in liquid level elevation in real time and the side pressure dissipation characteristics are fitted by a convolutional time series model to achieve adaptive adjustment of the pumping stroke frequency, ensuring pouring safety. A two-dimensional deviation distribution matrix is ​​constructed by collecting the three-dimensional point cloud of the wall. The dominant frequency is analyzed by frequency domain transformation and the spacing of the secondary keel is matched to accurately locate stiffness failure components and back-calculate the maximum allowable support span. A full-process digital closed-loop control mechanism is constructed to eliminate blind spots in manual monitoring and improve the construction accuracy and safety management efficiency of the main structure. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a schematic diagram of the steps of the present invention;

[0048] Figure 2 This is a detailed schematic diagram of S1 of the present invention;

[0049] Figure 3 This is a detailed schematic diagram of S2 of the present invention;

[0050] Figure 4 This is a detailed schematic diagram of S3 of the present invention;

[0051] Figure 5 This is a detailed schematic diagram of S4 of the present invention;

[0052] Figure 6 This is a detailed schematic diagram of S5 of the present invention;

[0053] Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0054] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0055] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0056] Please see Figure 1 This invention provides an optimization method for the construction of building main structures based on AI visual recognition, including the following steps:

[0057] S1: Acquire video frame sequence of the object being lifted by the tower crane, calculate pixel displacement to generate the centroid displacement vector of the object being lifted, construct a grid space based on the velocity direction, calculate the diffusion radius and probability based on the speed and direction of the object being lifted and map them to generate a time-varying probability grid map.

[0058] S2: Obtain the planned path of the concrete pump truck boom, map it to the time-varying probability grid map for collision risk assessment, respond to the assessment results and drive the camera to acquire depth map, segment the liquid surface area and calculate the height difference, and aggregate to generate an incremental sequence of liquid surface vertical elevation.

[0059] S3: Input the vertical elevation increment sequence of the liquid surface into the convolutional time series processing structure, use the exponential decay weight to fit the side pressure dissipation characteristics, obtain the equivalent side pressure index by weighted summation of historical data, compare the equivalent side pressure index with the bearing standard, and generate the pumping stroke frequency adjustment signal.

[0060] S4: In response to the termination of the transmission of the pumping stroke frequency adjustment signal, collect three-dimensional point cloud data of the wall surface, fit and calculate the deviation angle between the normal vector and the design normal vector, map it to two-dimensional plane coordinates, and construct a two-dimensional deviation distribution matrix.

[0061] S5: Extract the dominant frequency components from the two-dimensional deviation distribution matrix through frequency domain transformation, convert them into physical wavelengths and compare them with the spacing of the secondary keel to identify failed components. Calculate the maximum allowable spacing configuration data of the support components based on the elastic modulus and generate structural construction record data.

[0062] The time-varying probability grid map includes grid coordinate index, spatial occupancy probability value, and dynamic diffusion boundary marker; the liquid surface vertical elevation increment sequence includes sampling timestamp, liquid surface vertical height difference, and pouring area marker; the pumping stroke frequency adjustment signal includes stroke frequency target value, pressure safety margin coefficient, and pumping rate control command code; the two-dimensional deviation distribution matrix includes plane mapping coordinates, normal deviation angle value, and local flatness level marker; and the structural construction record data includes stiffness failure component location index, optimized secondary keel spacing parameters, and frequency domain feature verification results.

[0063] Please see Figure 2 The specific steps for obtaining a time-varying probability raster map are as follows:

[0064] S101: Acquire the video frame sequence of the tower crane hoisting object recorded by the visual acquisition device, call the preset camera intrinsic parameter matrix and spatial calibration coefficient, detect and lock the outline region of the hoisting object in the video frame, extract the pixel coordinates of the geometric center of the outline region, calculate the pixel coordinate difference of the geometric center between the current frame and the previous frame, and combine the frame rate parameter of the acquisition device to convert the pixel coordinate difference into a velocity vector to generate the displacement vector of the hoisting object's center of gravity.

[0065] Video frame sequences of the crane's lifting area are recorded using an industrial-grade high-definition vision acquisition device mounted on top of the crane. This device employs a global shutter CMOS sensor with a resolution of 4096 pixels multiplied by 2160 pixels and a frame rate of 60 frames per second. While acquiring the video stream data, it simultaneously calls a preset camera intrinsic parameter matrix and spatial calibration coefficients. Specifically, the camera intrinsic parameter matrix is ​​constructed by acquiring a sequence of 20 multi-angle static images containing a 12x9 checkerboard calibration board. For each image, the spatial gradient magnitude of pixel grayscale in the horizontal and vertical directions is calculated; for example, the Sobel operator is used to calculate the grayscale in the X and Y directions respectively. The derivative is used to select pixel locations where the gradient magnitude simultaneously exhibits local maxima in orthogonal directions as calibration corner points. Subpixel-level interpolation is then performed to obtain corner point coordinates accurate to 0.1 pixels. Subsequently, a geometric mapping equation describing the projection relationship from 3D spatial coordinates to 2D image plane coordinates is established. The actual physical coordinates of the calibration corner points are substituted into the equation to construct a reprojection residual function characterizing the deviation between the calculated projection point and the actual corner point position. The Levenberg-Marquardt algorithm is used to update the focal length parameters, principal point coordinate parameters, and radial distortion coefficients through 100 iterations until the value of the reprojection residual function is minimized, for example, when the average reprojection error is minimized. The iteration stops when the value decreases to 0.3 pixels, thus constructing a camera intrinsic parameter matrix containing focal lengths fx and fy of 3500 pixels and principal point coordinates cx of 2048 pixels and cy of 1080 pixels. Based on this, a rectangular reference object with a known physical side length is identified within the working plane, such as a standard rectangular steel plate with a side length of 2000 mm. The coordinates of the four vertices of this rectangular reference object on the image plane are extracted. Distortion correction is performed on the vertex coordinates based on the radial distortion coefficients k1 and k2 in the camera intrinsic parameter matrix. The corrected pixel Euclidean distance between adjacent vertices is calculated; for example, if the calculated pixel distance is 400 pixels, then... The ratio of the known physical side length of 2000 mm to the pixel Euclidean distance of 400 pixels is calculated and established as the conversion benchmark between the image pixel unit of 5 mm per pixel and the physical space length, generating a spatial calibration coefficient. The above parameters are used to detect and lock the outline region of the suspended object in the video frame. For example, a deep neural network model based on YOLOv8 is used to extract the pixel coordinates of the geometric center of the outline region. The pixel coordinate difference between the geometric center coordinates of the current frame (e.g., 1024, 500) and the geometric center coordinates of the previous frame (e.g., 1020, 497) is calculated. Combined with the frame rate parameter of the acquisition device, the pixel coordinate difference is converted into a velocity vector to generate the displacement vector of the suspended object's center of gravity.

[0066] S102: Call the displacement vector of the suspended object's center of gravity, analyze the vector magnitude as the instantaneous motion rate of the suspended object, analyze the vector angle as the motion direction angle, construct a two-dimensional grid coordinate system with the current position of the suspended object as the origin, calculate the risk diffusion distance based on the instantaneous motion rate, determine the angular range of the diffusion fan area in combination with the motion direction angle, use the risk diffusion distance as the radial length limit of the fan area, and generate the radius of the grid diffusion area.

[0067] The displacement vector of the suspended object's center of gravity is called, and the vector magnitude is analyzed as the instantaneous velocity of the suspended object. The vector angle is analyzed as the direction angle of motion. For example, when the horizontal displacement component is 4 pixels (i.e., 1024 minus 1020) and the vertical displacement component is 3 pixels (i.e., 500 minus 497), the calculated vector magnitude is 5 pixels. Combining the aforementioned calibration coefficient of 5 millimeters per pixel and a frame interval of 1 / 60 second, the instantaneous velocity is calculated to be 1.5 meters per second, and the direction angle is 37 degrees. Then, a two-dimensional grid coordinate system is constructed with the current position of the suspended object as the origin, and the grid resolution is set to 0.1 meters multiplied by 0.1 meters. Based on the instantaneous velocity... The risk diffusion distance is calculated based on the motion rate. This calculation process uses a linear prediction model, multiplying the instantaneous motion rate by a preset system reaction time, such as 2 seconds, and adding a safety buffer distance, such as 1 meter, to calculate a risk diffusion distance of 4 meters. The angular range of the diffusion fan-shaped area is determined by combining the motion direction angle. For example, a range of 30 degrees to the left and right of the motion direction angle is set. The risk diffusion distance of 4 meters is used as the radial length limit of the fan-shaped area to generate the radius of the grid diffusion area. This radius defines a fan-shaped potential collision area with the suspended object as the center, a radius of 4 meters, and an opening facing the motion direction.

[0068] S103: For each grid cell within the radius coverage of the grid diffusion area, obtain the grid center point coordinates, calculate the Euclidean distance between the grid center point coordinates and the predicted position of the suspended object, and map it as a collision risk weight. Normalize the collision risk weight to obtain the occupancy probability value, establish the correspondence between the occupancy probability value and the grid coordinate index, and generate a time-varying probability grid map.

[0069] The predicted position of the suspended object is obtained by extending the current position along the direction of motion by one time step. This Euclidean distance is input into a Gaussian decay function and mapped to a collision risk weight. This Gaussian decay function is centered at the predicted position of the suspended object and has a standard deviation of one-quarter of the risk diffusion distance. For example, for a grid cell 1 meter away from the predicted position, the calculated unnormalized risk weight is 0.6 (corresponding to e to the power of -0.5), while for a grid cell 3 meters away, the weight decays to 0.01 (corresponding to e to the power of -4.5). The calculated collision risk weight is then applied... The normalization process involves summing the weights of all grid cells within the statistical area, then dividing the weight of each individual grid cell by this sum to obtain an occupancy probability value between 0 and 1. For example, the occupancy probability value of a core area grid cell is calculated to be 0.85, while that of an edge area grid cell is 0.05. Finally, a correspondence is established between the occupancy probability value and the grid coordinate index, such as (row 50, column 60), to generate a time-varying probability grid map containing dynamic occupancy information. The real-time refresh rate of this map is consistent with the video capture frame rate. Table 1 shows the probability mapping data of some grid cells.

[0070] Table 1. Local Time-Varying Probabilistic Raster Map Data Table

[0071]

[0072] As shown in Table 1, the closer the grid is to the center of the suspended object, the higher its probability of being occupied, which intuitively reflects the dynamic risk distribution.

[0073] Please see Figure 3 The specific steps for obtaining the vertical elevation increment sequence of the liquid surface are as follows:

[0074] S201: Obtain the coordinate set of the planned path of the concrete pump truck boom, project and map the three-dimensional coordinate points in the planned path coordinate set to the two-dimensional index of the time-varying probability grid map, retrieve the space occupancy probability value in the corresponding grid, conduct a collision risk assessment, and output a safe operation permit instruction based on the assessment results.

[0075] The planned path coordinates of the concrete pump truck boom are obtained through a wireless communication interface. This set contains a series of three-dimensional coordinate points (X, Y, Z) arranged in chronological order. First, the three-dimensional coordinate points in the planned path coordinate set are projected and mapped onto the two-dimensional index of a time-varying probability raster map. Specifically, this is achieved by ignoring the Z-axis height information and dividing the X and Y coordinates by the raster resolution of 0.1 meters and rounding them down. For example, the path point (10.5 meters, 12.3 meters, 20 meters) is mapped to the raster index (105, 123). Then, the corresponding raster index is retrieved in the current time-varying probability raster map. The space occupancy probability values ​​in the diagram are used to perform a collision risk assessment. This assessment process sets a safety threshold, such as 0.3. If the grid occupancy probability value mapped to any point on the path exceeds 0.3, a collision risk is determined. Conversely, if the probability values ​​corresponding to all path points are below 0.3, for example, the probability of path point A is 0.05 and the probability of path point B is 0.10, then it is determined to be safe. Based on the assessment results, a safe operation permit instruction is output. This instruction is a binary signal that is only set to a high level of 1 when the safety verification of the entire path is passed, thereby allowing subsequent operations.

[0076] S202: In response to the safety work permit instruction, the depth camera is activated to acquire surface depth images of the liquid concrete being poured. The spatial gradient magnitude of each pixel in the depth image is calculated through edge detection. Pixels with spatial gradient magnitudes lower than the flatness threshold are filtered out and aggregated into connected components. The depth measurement values ​​of all pixels in the connected components are extracted. The association mapping between the depth measurement values ​​and the acquisition time is established to generate a time-segmented liquid surface depth data set.

[0077] In response to a high-level signal from a safety work permit command, a ToF depth camera installed at the end of the pump truck boom is activated. The acquisition frequency is set to 10 Hz to capture surface depth images of the molten concrete being poured. The spatial gradient amplitude of each pixel in the depth image is calculated using the Canny edge detection operator. Pixels with spatial gradient amplitudes below a flatness threshold are filtered out. This flatness threshold is determined during the initialization phase before the monitoring task begins: a stationary, horizontal concrete surface within the construction site is selected as a reference object, a baseline depth image of the reference object is acquired, and the spatial gradient amplitude of each pixel in the baseline depth image is calculated point-by-point. All pixels with spatial gradient amplitudes below a flatness threshold are then filtered out. Statistical analysis is performed on the gradient amplitude samples to construct a Gaussian normal distribution model. The expected value of the model is extracted, for example, 2 mm per pixel, and the standard deviation parameter is extracted, for example, 0.5 mm per pixel. The sum of the expected value and three times the standard deviation is calculated, i.e., 2 plus 1.5 equals 3.5 mm per pixel, which is set as the flatness threshold covering the inherent measurement noise range of the sensor. In real-time monitoring, pixels with gradient amplitudes lower than 3.5 mm per pixel are identified as flat regions, and they are aggregated into connected components using a flooding filling algorithm. The depth measurement values ​​of all pixels in the connected components are extracted, and a correlation mapping between the depth measurement values ​​and the acquisition time is established to generate a time-segmented liquid surface depth data set.

[0078] S203: Call the time-segmented liquid surface depth data set, calculate the arithmetic mean of the pixel depth measurement values ​​in the connected domain at each acquisition time as the instantaneous liquid surface height at the target time, select two adjacent instantaneous liquid surface heights in the time series to perform numerical difference calculation, obtain the vertical height change within a unit time step, reorganize the vertical height change according to the time order, and generate a liquid surface vertical elevation increment sequence.

[0079] The system calls upon a time-segmented liquid surface depth data set and calculates the arithmetic mean of pixel depth measurements within the connected domain at each acquisition time as the instantaneous liquid surface height at the target time. It then selects two adjacent instantaneous liquid surface heights in the time series and performs numerical difference calculations. For example, if the liquid surface height at time T2 is 5.05 meters and at time T1 is 5.00 meters, with a time step of 1 minute, the difference of 0.05 meters is calculated as the vertical height change per unit time step. This change reflects the rate of concrete pouring (approximately 3 meters per hour). The continuously calculated vertical height changes are then rearranged in chronological order to construct a one-dimensional time series vector, generating a sequence of incremental vertical elevations of the liquid surface. This sequence clearly records the historical thickness of each layer of concrete during the pouring process.

[0080] Please see Figure 4 The specific steps for obtaining the pump stroke frequency adjustment signal are as follows:

[0081] S301: Obtain the initial setting time parameter and hardening rate coefficient of concrete material, construct a numerical decay function that decreases exponentially with time based on the hardening rate coefficient, calculate the signal retention weight value corresponding to each historical moment based on the sampling time step, construct a time-series processing structure containing multi-layer causal convolution operation units, map the signal retention weight value to the weight parameter matrix of the convolution kernel, and generate a time-varying decay convolution kernel for lateral pressure.

[0082] The initial setting time parameter of the current batch of concrete material is obtained from the material database, for example, 4 hours (240 minutes), and the hardening rate coefficient is obtained, for example, 0.01. Based on the hardening rate coefficient, a numerical decay function is constructed that decreases exponentially over time. This function characterizes the physical property that the contribution of freshly poured concrete to the lateral pressure of the formwork decreases over time. Specifically, the function form is that the weight value is equal to the negative power of the natural constant e, and the exponent is the hardening rate coefficient multiplied by the time interval. Based on the sampling time step, for example, 1 minute, the signal retention weight value corresponding to each historical moment is calculated. For example, for the current... The concrete layer poured 10 minutes before the time step is calculated with a weight of e to the power of -0.1, which is approximately 0.904. For the layer poured 100 minutes before the time step, the weight decays to e to the power of -1, which is approximately 0.367. A time-series processing structure containing multiple causal convolution operation units is constructed. This structure includes an input layer, three dilated convolution layers and an output layer. The calculated signal retains the weight value sequence, is flipped and normalized, and then directly mapped and assigned to the weight parameter matrix of the convolution kernel to generate a time-varying decay convolution kernel for lateral pressure. This ensures that the convolution operation can physically simulate the time dissipation process of pressure.

[0083] S302: Call the incremental sequence of vertical elevation of liquid surface into the time series processing structure, use the time-varying decay convolution kernel of lateral pressure to perform sliding window convolution operation on the sequence data, multiply each historical incremental data point covered by the window by the corresponding weight parameter, perform cumulative calculation on all weighted data points, analyze the lateral pressure superposition effect, output the pressure estimate value at the current moment, and generate the equivalent lateral pressure index.

[0084] The sliding window length is set to cover 240 data points over the past 4 hours (in 1-minute increments). For each historical incremental data point covered by the window, the corresponding weight parameter is multiplied. For example, for an increment of 0.05 meters at time T minus 10 in history, it is multiplied by the corresponding weight 0.905 (e to the power of -0.1), resulting in a weighted contribution value of 0.04525 meters, which is equivalent to the liquid column height. The weighted data points within the window are summed. This summation result represents the total effective liquid column height still acting on the template at the current moment after considering hardening dissipation. Then, this total effective liquid column height is multiplied by the density of concrete, for example, 24 kN per cubic meter, to analyze the lateral pressure superposition effect and output the pressure estimate value at the current moment. For example, if the effective liquid column height is calculated to be 2.5 meters, then the lateral pressure is 60 kN per square meter, generating the equivalent lateral pressure index. Table 2 shows the data changes during the convolution calculation process.

[0085] Table 2 Implementation Data Table for Lateral Pressure Convolution Calculation

[0086]

[0087] As shown in Table 2, although the initial increment is the same, the contribution of early concrete to the current lateral pressure decreases significantly over time.

[0088] S303: Based on the equivalent lateral pressure index, retrieve and compare the preset standard values ​​of the ultimate bending capacity of the support components, calculate the numerical deviation as a safety margin index, adjust the target operating frequency of the pumping stroke, convert the target operating frequency into equipment control commands, and generate a pumping stroke frequency adjustment signal.

[0089] Based on an equivalent lateral pressure index, such as 60 kN / m², the preset standard values ​​of the ultimate bending capacity of the supporting components are retrieved and compared. For example, if the maximum allowable lateral pressure of the formwork system is set to 80 kN / m², the numerical deviation is calculated as a safety margin index, i.e., 80 minus 60 equals 20 kN / m². Based on this safety margin index, the target operating frequency of the pumping stroke is adjusted using proportional control logic. When the safety margin is large (e.g., greater than 15), the frequency is allowed to be increased to speed up construction. When the margin is small (e.g., less than 5), the frequency is forcibly reduced. For example, if the base frequency is set to 20 times per minute and the gain coefficient is 0.5, the adjustment amount is 20 multiplied by 0.5 equals 10, and the target frequency is increased to 30 times per minute. The target operating frequency is converted into equipment control instructions, i.e., a hexadecimal control message containing a specific frequency code is generated, and a pumping stroke frequency adjustment signal is generated and sent to the pump truck controller.

[0090] Please see Figure 5 The specific steps for obtaining the two-dimensional deviation distribution matrix are as follows:

[0091] S401: Real-time transmission status monitoring is performed on the pumping stroke frequency adjustment signal. When the pumping operation ends, the laser scanner is triggered to perform a panoramic ranging scan on the completed wall surface, collect laser echo data and convert it into spatial three-dimensional coordinates, remove outlier noise, establish the index relationship between three-dimensional coordinate points and scanning intensity, and generate a point cloud set of the wall surface to be measured.

[0092] Real-time transmission status monitoring is performed on the pumping stroke frequency adjustment signal. When the signal contains a status code of "stop pumping" or "operation completed", the ground-based 3D laser scanner set up at the construction site is triggered in response to the transmission termination status to perform a panoramic ranging scan on the completed wall surface. The scanning resolution is set to 6 mm @ 10 m. Laser echo data is collected and converted into spatial 3D coordinates (X, Y, Z). Using a statistical outlier removal (SOR) filter, with the number of neighborhood points set to 50 and the standard deviation multiple set to 1.0, outlier noise points caused by air dust or edge scattering are removed. An index relationship between 3D coordinate points and scanning intensity is established to generate a point cloud set of the wall surface to be measured containing millions of high-precision coordinate points.

[0093] S402: For the three-dimensional coordinate points in the point cloud set of the wall surface to be measured, construct a spherical neighborhood with the coordinate points as the center and retrieve neighboring points. Construct the covariance matrix of the local plane based on the least squares method, analyze the eigenvalues ​​and eigenvectors of the matrix, select the eigenvector corresponding to the smallest eigenvalue as the normal direction, perform vector magnitude normalization operation, establish the mapping relationship between coordinate points and normal vectors, and generate the measured local unit normal vector set;

[0094] For each 3D coordinate point in the point cloud set of the wall surface to be measured, a spherical neighborhood with a radius of 50 mm is constructed with that coordinate point as the center, and all neighboring points within the neighborhood are retrieved, for example, 20 neighboring points are retrieved. The covariance matrix of the local plane is constructed based on the least squares method. Specifically, the coordinates of all points in the neighborhood are subtracted from the centroid coordinates to form matrix A. The transpose of A is calculated and multiplied by A to obtain a 3x3 covariance matrix. The three eigenvalues ​​(λ1, λ2, λ3) of this matrix and their corresponding eigenvectors are analyzed. The eigenvector corresponding to the smallest eigenvalue is selected as the normal direction of the local plane. This direction represents the perpendicular direction of the surface at that point. The vector magnitude normalization operation is performed. The three components of the eigenvector are divided by their magnitudes to make their lengths 1, and the mapping relationship between the coordinate points and the normal vectors is established to generate the measured local unit normal vector set.

[0095] S403: Call the measured local unit normal vector set, calculate the cosine value of the angle between each measured normal vector and the ideal design plane normal vector, obtain the angle deviation value through inverse cosine operation, extract the horizontal and vertical components of the three-dimensional coordinate points on the wall facade, map them into the row and column numbers of the two-dimensional grid, fill the angle deviation value into the grid, and generate a two-dimensional deviation distribution matrix.

[0096] The measured local unit normal vector set is called to obtain the ideal design plane normal vector. For example, for a vertical wall, the ideal normal vector is (0, 0, 1). The dot product of each measured normal vector and the ideal design plane normal vector is calculated to obtain the cosine value of the angle. For example, if the measured vector is (0.017, 0, 0.999), the dot product is 0.999. The angle deviation value is obtained by using the inverse cosine function (acos), and the deviation is calculated to be 0.045 radians (approximately 2.6 degrees). The horizontal (X-axis) and vertical (Y-axis) components of the three-dimensional coordinate points on the wall facade are extracted. The two-dimensional grid size is set to 100 mm x 100 mm. The X and Y components are divided by the grid size and rounded to the nearest integer, mapping them to the row and column numbers of the two-dimensional grid. For example, coordinates (2500, 3000) are mapped to (25, 1). 30) Fill the calculated angle deviation values ​​into the corresponding grid cells. If there are multiple points in a grid, take the average value to generate a two-dimensional deviation distribution matrix. This matrix intuitively transforms the three-dimensional flatness problem of the wall into a two-dimensional numerical heat map.

[0097] Please see Figure 6 The specific steps for obtaining structural construction record data are as follows:

[0098] S501: Call the two-dimensional deviation distribution matrix to perform eigentransformation from the spatial domain to the frequency domain, calculate the modulus square of the complex spectrum, construct the power spectral density matrix, lock the dominant frequency component in the energy concentration, calculate the reciprocal of the dominant frequency component based on the spatial domain sampling frequency, obtain the corresponding spatial period length, and generate the dominant deformation physical wavelength.

[0099] The two-dimensional deviation distribution matrix is ​​used to perform a feature transformation from the spatial domain to the frequency domain. Specifically, the two-dimensional fast Fourier transform (2D FFT) algorithm is used to calculate the sum of the squares of the real and imaginary parts of the transformed complex spectrum to obtain the power spectral density matrix. This matrix reveals the energy distribution of the wavy deformation on the wall surface. Peaks are searched in the power spectrum to identify the dominant frequency components where energy is concentrated. For example, a peak frequency f of 0.5 Hz / m is found on the horizontal frequency axis. Based on the spatial domain sampling frequency (i.e., the reciprocal of the grid spacing, 1 / 0.1 m = 10 Hz / m), the reciprocal of the dominant frequency component is calculated to obtain the corresponding spatial period length. The calculation process is 1 divided by 0.5 equals 2 meters, which indicates that there is a periodic undulation with a wavelength of 2 meters on the wall surface, generating the physical wavelength of the dominant deformation.

[0100] S502: Obtain the secondary keel design spacing parameters in the architectural design drawings, perform numerical matching verification between the dominant deformation physical wavelength and the spacing parameters, analyze the correlation between the wall deformation cycle and the skeleton arrangement, determine the type of support component that affects deformation, lock the target support component with stiffness failure, and generate a stiffness failure component location index.

[0101] Obtain the secondary keel design spacing parameters from the architectural design drawings. For example, the horizontal spacing design value of the secondary keel (supporting square steel) is 0.3 meters. Verify the numerical matching between the dominant deformation physical wavelength of 2 meters and the spacing parameter of 0.3 meters. Specifically, calculate the ratio between the two or find the multiple relationship. If the physical wavelength (2 meters) is a non-integer multiple of the spacing (0.3 meters) and highly matches the spacing of the main keel (e.g., 2 meters), analyze the correlation between the wall deformation cycle and the skeleton arrangement to determine the type of supporting component affecting the deformation. In this example, the component that fails in stiffness is determined to be the main keel rather than the secondary keel because the deformation wavelength is consistent with the spacing of the main keel. Conversely, if the wavelength is 0.6 meters, which is exactly twice the spacing of the secondary keel, it is determined to be double-span deformation caused by the excessive span of the secondary keel. Based on this logic, lock the target supporting component that fails in stiffness and generate a location index for the stiffness-failed component.

[0102] S503: For the target component of the stiffness failure component location index, obtain the elastic modulus constant of the corresponding material and the target flatness requirement, calculate the maximum allowable spacing configuration data of the support components, integrate the stiffness failure analysis results and the spacing configuration data, and generate structural construction record data;

[0103] For the target component (e.g., secondary keel) identified by the stiffness failure component positioning index, the elastic modulus constant E of the corresponding material (e.g., Q235 steel, which is 206 gigapascals (GPa)) and the target flatness requirement (e.g., the maximum allowable deflection w_max = L / 400) are obtained. The deformation formula of the Euler-Bernoulli beam theory is used for inverse calculation. The formula logic is that the maximum deflection is directly proportional to the load, directly proportional to the fourth power of the spacing, and inversely proportional to the elastic modulus and moment of inertia. By fixing the load and material properties, the maximum spacing L that meets the flatness requirement is solved. For example, if the currently calculated theoretical maximum allowable spacing is 0.25 meters, while the actual design spacing is 0.3 meters, it indicates that the original design stiffness is insufficient, and the maximum allowable spacing configuration data for the support components is calculated to be 0.25 meters. Finally, the stiffness failure analysis results and spacing configuration data are integrated to generate structural construction record data, as shown in Table 3, which details the diagnosis and optimization results.

[0104] Table 3 Structural Construction Stiffness Diagnosis Record Form

[0105]

[0106] As shown in Table 3, the system clearly identified the risk of insufficient rigidity in the main keel and provided specific suggestions for reducing the spacing. This data was directly used to guide subsequent construction and rectification.

[0107] Please see Figure 7 A construction method optimization system for building main structures based on AI visual recognition, including:

[0108] The dynamic risk raster mapping module acquires video frame sequences of objects lifted by tower cranes, calculates pixel displacement to generate the centroid displacement vector of the object, constructs a raster space based on the velocity direction, calculates and maps the diffusion radius and probability based on the speed and direction of the object's movement, and generates a time-varying probability raster map.

[0109] The liquid surface increment sequence extraction module obtains the planned path of the concrete pump truck arm, maps it to a time-varying probability grid map for collision risk assessment, responds to the assessment results and drives the camera to acquire depth maps, segments the liquid surface area and calculates the elevation difference, and aggregates to generate a liquid surface vertical elevation increment sequence.

[0110] The pumping frequency feedback control module inputs the vertical elevation increment sequence of the liquid surface into the convolutional time series processing structure, uses exponential decay weights to fit the side pressure dissipation characteristics, obtains the equivalent side pressure index by weighted summation of historical data, compares the equivalent side pressure index with the bearing standard, and generates a pumping stroke frequency adjustment signal.

[0111] The two-dimensional deviation matrix construction module, in response to the termination state of the transmission of the pumping stroke frequency adjustment signal, collects three-dimensional point cloud data of the wall surface, fits and calculates the deviation angle between the normal vector and the design normal vector, maps it to two-dimensional plane coordinates, and constructs a two-dimensional deviation distribution matrix.

[0112] The structural stiffness diagnosis and recording module extracts the dominant frequency components from the two-dimensional deviation distribution matrix through frequency domain transformation, converts them into physical wavelengths, compares them with the spacing of the secondary keel, identifies failed components, calculates the maximum allowable spacing configuration data of the support components based on the elastic modulus, and generates structural construction record data.

[0113] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for optimizing the construction techniques of building main structures based on AI visual recognition, characterized in that: Includes the following steps: S1: Acquire video frame sequence of the object being lifted by the tower crane, calculate pixel displacement to generate the centroid displacement vector of the object being lifted, construct a grid space based on the velocity direction, calculate the diffusion radius and probability based on the speed and direction of the object being lifted and map them to generate a time-varying probability grid map. S2: Obtain the planned path of the concrete pump truck boom, map it to the time-varying probability grid map to perform a collision risk assessment, respond to the assessment results and drive the camera to acquire a depth map, segment the liquid surface area and calculate the height difference, and aggregate to generate an incremental sequence of vertical elevation of the liquid surface. S3: Input the vertical elevation increment sequence of the liquid surface into the convolutional time series processing structure, use the exponential decay weight to fit the side pressure dissipation characteristics, obtain the equivalent side pressure index by weighted summation of historical data, compare the equivalent side pressure index with the bearing standard, and generate the pumping stroke frequency adjustment signal. S4: In response to the termination of the transmission of the pumping stroke frequency adjustment signal, collect three-dimensional point cloud data of the wall surface, fit and calculate the deviation angle between the normal vector and the design normal vector, map it to two-dimensional plane coordinates, and construct a two-dimensional deviation distribution matrix. S5: Extract the dominant frequency component from the frequency domain transformation of the two-dimensional deviation distribution matrix, convert it into physical wavelength and compare it with the spacing of the secondary keel, identify the failed components, calculate the maximum allowable spacing configuration data of the support components based on the elastic modulus, and generate structural construction record data.

2. The method for optimizing the construction method of building main structure based on AI visual recognition according to claim 1, characterized in that, The time-varying probability grid map includes grid coordinate index, spatial occupancy probability value, and dynamic diffusion boundary marker; the liquid surface vertical elevation increment sequence includes sampling timestamp, liquid surface vertical height difference, and pouring area marker; the pumping stroke frequency adjustment signal includes stroke frequency target value, pressure safety margin coefficient, and pumping rate control command code; the two-dimensional deviation distribution matrix includes plane mapping coordinates, normal deviation angle value, and local flatness level marker; and the structural construction record data includes stiffness failure component positioning index, optimized secondary keel spacing parameters, and frequency domain feature verification results.

3. The method for optimizing the construction method of building main structure based on AI visual recognition according to claim 1, characterized in that, The specific steps for obtaining the time-varying probability raster map are as follows: S101: Acquire the video frame sequence of the tower crane hoisting object recorded by the visual acquisition device, call the preset camera intrinsic parameter matrix and spatial calibration coefficient, detect and lock the outline region of the hoisting object in the video frame, extract the pixel coordinates of the geometric center of the outline region, calculate the pixel coordinate difference of the geometric center between the current frame and the previous frame, and combine the frame rate parameter of the acquisition device to convert the pixel coordinate difference into a velocity vector to generate the displacement vector of the hoisting object's center of gravity. S102: Call the displacement vector of the suspended object's center of gravity, analyze the vector magnitude as the instantaneous motion rate of the suspended object, analyze the vector angle as the motion direction angle, construct a two-dimensional grid coordinate system with the current position of the suspended object as the origin, calculate the risk diffusion distance based on the instantaneous motion rate, determine the angular range of the diffusion fan-shaped area in combination with the motion direction angle, use the risk diffusion distance as the radial length limit of the fan-shaped area, and generate the radius of the grid diffusion area. S103: For each grid cell within the radius coverage of the grid diffusion area, obtain the grid center point coordinates, calculate the Euclidean distance between the grid center point coordinates and the predicted position of the suspended object, and map it as a collision risk weight. Normalize the collision risk weight to obtain the occupancy probability value, establish the correspondence between the occupancy probability value and the grid coordinate index, and generate a time-varying probability grid map.

4. The method for optimizing the construction method of building main structure based on AI visual recognition according to claim 3, characterized in that, The process of calling the preset camera intrinsic parameter matrix and spatial calibration coefficients specifically includes: Collect a multi-angle static image sequence containing a checkerboard calibration board, calculate the spatial gradient magnitude of the image pixel grayscale in the horizontal and vertical directions, select the pixel positions where the gradient magnitude simultaneously presents a local maximum in the orthogonal direction as calibration corner points, and perform sub-pixel level interpolation positioning. A geometric mapping equation describing the projection relationship from three-dimensional spatial coordinates to two-dimensional image plane coordinates is established. The actual coordinates of the calibrated corner points are substituted into the equation to construct a reprojection residual function that characterizes the deviation between the calculated projection point and the actual corner point position. The focal length, principal point coordinates, and radial distortion coefficient are updated through multiple iterations to minimize the value of the reprojection residual function and construct the camera intrinsic parameter matrix. Identify a rectangular reference object with a known physical side length within the working plane, extract the vertex coordinates of the rectangular reference object on the image plane, perform distortion correction on the vertex coordinates based on the radial distortion coefficient in the camera intrinsic parameter matrix, calculate the pixel Euclidean distance between adjacent vertices after correction, calculate the ratio of the known physical side length to the pixel Euclidean distance, establish it as the conversion benchmark between image pixel units and physical space length, and generate the spatial calibration coefficient.

5. The method for optimizing the construction method of building main structure based on AI visual recognition according to claim 3, characterized in that, The specific steps for obtaining the vertical elevation increment sequence of the liquid surface are as follows: S201: Obtain the set of coordinates for the planned path of the concrete pump truck boom, project and map the three-dimensional coordinate points in the set of planned path coordinates onto the two-dimensional index of the time-varying probability grid map, retrieve the space occupancy probability value in the corresponding grid, conduct a collision risk assessment, and output a safe operation permit instruction based on the assessment results. S202: In response to the safety operation permit instruction, start the depth camera, acquire the surface depth image of the liquid concrete being poured, calculate the spatial gradient amplitude of each pixel in the depth image through edge detection, filter the pixels with spatial gradient amplitude below the flatness threshold and aggregate them into connected components, extract the depth measurement values ​​of all pixels in the connected components, establish the association mapping between the depth measurement values ​​and the acquisition time, and generate a time-segmented liquid surface depth data set. S203: Call the time-segmented liquid surface depth data set, calculate the arithmetic mean of the pixel depth measurement values ​​in the connected domain at each acquisition time as the instantaneous liquid surface height at the target time, select two adjacent instantaneous liquid surface heights in the time series to perform numerical difference calculation, obtain the vertical height change within a unit time step, reorganize the vertical height change according to the time order, and generate a liquid surface vertical elevation increment sequence.

6. The method for optimizing the construction method of building main structure based on AI visual recognition according to claim 5, characterized in that, The process of filtering pixels whose spatial gradient magnitude is lower than the flatness threshold specifically includes: In the initialization phase before the monitoring task is started, a stationary horizontal plane in the construction site is selected as the reference object. The reference depth image of the reference object is collected, and the spatial gradient amplitude of each pixel in the reference depth image is calculated point by point. Statistical analysis is performed on all spatial gradient amplitude samples to construct a Gaussian normal distribution model. The mathematical expectation and standard deviation parameters of the model are extracted, and the sum of the mathematical expectation and three times the standard deviation is calculated and set as the flatness threshold that covers the inherent measurement noise range of the sensor. During real-time monitoring, the pixel spatial gradient amplitude of the depth image of the surface of the liquid concrete being poured is calculated. The amplitude is compared with the flatness threshold. Pixels with spatial gradient amplitudes lower than the flatness threshold are selected, and the pixels are determined to belong to the flat fluid surface area. They are marked as valid liquid surface data points and the corresponding depth measurement values ​​are retained.

7. The method for optimizing the construction method of building main structure based on AI visual recognition according to claim 5, characterized in that, The specific steps for obtaining the pumping stroke frequency adjustment signal are as follows: S301: Obtain the initial setting time parameter and hardening rate coefficient of concrete material, construct a numerical decay function that decreases exponentially with time based on the hardening rate coefficient, calculate the signal retention weight value corresponding to each historical moment based on the sampling time step, construct a time-series processing structure containing multi-layer causal convolution operation units, map the signal retention weight value to the weight parameter matrix of the convolution kernel, and generate a time-varying decay convolution kernel for lateral pressure. S302: Call the liquid surface vertical elevation increment sequence into the time series processing structure, use the side pressure time-varying decay convolution kernel to perform sliding window convolution operation on the sequence data, multiply each historical increment data point covered by the window by the corresponding weight parameter, perform cumulative calculation on all weighted data points, analyze the side pressure superposition effect, output the pressure estimate value at the current moment, and generate the equivalent side pressure index. S303: Based on the equivalent lateral pressure index, retrieve and compare the preset standard values ​​of the ultimate bending capacity of the support member, calculate the numerical deviation as a safety margin index, adjust the target operating frequency of the pumping stroke, convert the target operating frequency into equipment control commands, and generate a pumping stroke frequency adjustment signal.

8. The method for optimizing the construction method of building main structure based on AI visual recognition according to claim 7, characterized in that, The specific steps for obtaining the two-dimensional deviation distribution matrix are as follows: S401: Real-time transmission status monitoring is performed on the pumping stroke frequency adjustment signal. When the pumping operation ends, the laser scanner is triggered to perform a panoramic ranging scan on the surface of the completed wall, collect laser echo data and convert it into three-dimensional spatial coordinates, remove outlier noise, establish an index relationship between the three-dimensional coordinate points and the scanning intensity, and generate a point cloud set of the wall surface to be measured. S402: For the three-dimensional coordinate points in the point cloud set of the wall surface to be measured, construct a spherical neighborhood with the coordinate points as the center and retrieve neighboring points. Construct the covariance matrix of the local plane based on the least squares method, analyze the eigenvalues ​​and eigenvectors of the matrix, select the eigenvector corresponding to the smallest eigenvalue as the normal direction, perform vector magnitude normalization operation, establish the mapping relationship between coordinate points and normal vectors, and generate the measured local unit normal vector set; S403: Call the measured local unit normal vector set, calculate the cosine value of the angle between each measured normal vector and the ideal design plane normal vector, obtain the angle deviation value through inverse cosine operation, extract the horizontal and vertical components of the three-dimensional coordinate points on the wall facade, map them into the row and column numbers of the two-dimensional grid, fill the angle deviation value into the grid, and generate a two-dimensional deviation distribution matrix.

9. The method for optimizing the construction method of building main structure based on AI visual recognition according to claim 8, characterized in that, The specific steps for obtaining the structural construction record data are as follows: S501: Call the two-dimensional deviation distribution matrix to perform a feature transformation from the spatial domain to the frequency domain, calculate the modulus square of the complex spectrum, construct the power spectral density matrix, lock the dominant frequency component in the energy concentration, calculate the reciprocal of the dominant frequency component based on the spatial domain sampling frequency, obtain the corresponding spatial period length, and generate the dominant deformation physical wavelength. S502: Obtain the secondary keel design spacing parameters in the architectural design drawings, perform numerical matching verification between the dominant deformation physical wavelength and the spacing parameters, analyze the correlation between the wall deformation cycle and the skeleton arrangement, determine the type of support component that affects the deformation, lock the target support component with stiffness failure, and generate a stiffness failure component location index. S503: For the target component of the stiffness failure component positioning index, obtain the elastic modulus constant of the corresponding material and the target flatness requirement, calculate the maximum allowable spacing configuration data of the support components, integrate the stiffness failure analysis results and the spacing configuration data, and generate structural construction record data.

10. A construction method optimization system for building main structures based on AI visual recognition, characterized in that, The system is used to implement the construction method optimization method for building main structure based on AI visual recognition as described in any one of claims 1-9, and the system includes: The dynamic risk raster mapping module acquires video frame sequences of objects lifted by tower cranes, calculates pixel displacement to generate the centroid displacement vector of the object, constructs a raster space based on the velocity direction, calculates and maps the diffusion radius and probability based on the speed and direction of the object's movement, and generates a time-varying probability raster map. The liquid surface increment sequence extraction module obtains the planned path of the concrete pump truck arm, maps it to the time-varying probability grid map for collision risk assessment, responds to the assessment result and drives the camera to acquire depth map, segments the liquid surface area and calculates the height difference, and aggregates to generate a liquid surface vertical elevation increment sequence. The pumping frequency feedback control module inputs the vertical elevation increment sequence of the liquid surface into the convolutional time series processing structure, uses exponential decay weights to fit the side pressure dissipation characteristics, obtains the equivalent side pressure index by weighted summation of historical data, compares the equivalent side pressure index with the bearing standard, and generates a pumping stroke frequency adjustment signal. The two-dimensional deviation matrix construction module, in response to the termination state of the transmission of the pumping stroke frequency adjustment signal, collects three-dimensional point cloud data of the wall surface, fits and calculates the deviation angle between the normal vector and the design normal vector, maps it to two-dimensional plane coordinates, and constructs a two-dimensional deviation distribution matrix. The structural stiffness diagnosis and recording module extracts the dominant frequency components from the frequency domain transformation of the two-dimensional deviation distribution matrix, converts them into physical wavelengths, compares them with the spacing of the secondary keel, identifies failed components, calculates the maximum allowable spacing configuration data of the support components based on the elastic modulus, and generates structural construction record data.

Images