Precast beam plate process intelligent detection method and system based on target detection

Abstract

This invention discloses an intelligent detection method and system for precast beam and slab production processes based on target detection. The method includes: deploying process cameras on the precast slab production line, selecting the closest point to the midpoint of the overhead view plane as the calibration reference, defining a world coordinate system with its optical center ground projection as the origin, and calibrating the deployed cameras. A training model is used to identify process targets, and a tracking method is used to associate continuous frames to form short-term trajectories, transforming the coordinates. When a target leaves the camera overlap area, a time threshold is set to search for a new trajectory, associate the trajectories, and assign a global ID. An intelligent data entry system hardware is constructed to determine the correspondence between process targets. Upon detection of "reinforcing cage hoisting," a virtual ID is generated; other processes are retrieved based on time, and data entry is completed and encrypted. A multi-dimensional matching strategy is implemented to achieve cross-camera trajectory association, and a digital twin-driven visualization of production progress is achieved, improving the intelligent management level of precast beam and slab production.

Description

Technical Field

[0001] This invention belongs to the field of precast beam and slab manufacturing in bridge engineering, and more specifically, relates to an intelligent detection method and system for precast beam and slab processes based on target detection. Background Technology

[0002] In the field of bridge construction, precast concrete bridge decks are widely used due to their advantages such as high construction efficiency and controllable quality. Currently, the progress control of precast slab production mainly relies on traditional manual data entry methods, requiring on-site personnel to use mobile software to upload information such as process, time, beam code, and platform code. However, for large-scale precast slab projects, construction personnel need to frequently enter data for multiple processes, which not only consumes a lot of energy but also easily leads to the loss of beam process information, making it difficult to meet the management requirements for real-time and complete process information during production.

[0003] In recent years, Radio Frequency Identification (RFID) technology has begun to be applied to the data entry of prefabricated structure processes. This technology binds the prefabricated components to the trolley by installing a sensor chip on its bottom, and two readers are installed at the entry and exit points of each work area. When the trolley reaches the first reader (starting position) in the work area, the reader automatically records the current time as the start time of the process for that component; when it reaches the second reader (leaving position), it records the end time of the process. The position of the trolley and the prefabricated component is then determined by analyzing the data transmitted from the readers. However, for large prefabricated slab projects, due to the numerous work areas and the high cost of RFID hardware, its practical application suffers from excessive cost.

[0004] While computer vision-based process recognition technology has seen some applications in engineering manufacturing, it is relatively rare in the field of bridge component prefabrication. Unlike precast beam production, precast slabs are produced in multiple pieces simultaneously, and the production area is large, making it impossible for a single camera to provide comprehensive coverage. Existing tracking methods such as SORT, DeepSORT, and ByteTrack cannot achieve cross-regional tracking of precast slabs. Furthermore, due to the uniform appearance of precast slabs, re-identification methods based on appearance features cannot be used, making the movement and tracking of precast slabs between different work areas a significant challenge.

[0005] Traditional manual data entry methods are inefficient and prone to errors. During the multi-stage flow of precast slabs, human intervention inevitably leads to oversights, resulting in inaccurate process information and affecting the assessment of production progress. While RFID technology achieves some degree of automated recording, its high hardware costs and complex maintenance limit its widespread application in large-scale precast slab projects. Furthermore, the limitations of cross-regional tracking technology make it impossible to accurately track the movement of precast slabs throughout the production site, hindering comprehensive monitoring and management of the production process.

[0006] To address the aforementioned technical issues, this invention proposes an intelligent detection method for precast beam slab processes based on target detection. This method uses computer vision technology to identify precast slab processes in real time, automatically filling in process information, and simultaneously solving the problem of tracking precast slabs across camera areas. This improves the automation and intelligence level of precast slab production management and provides an effective solution for the digital management of bridge precast component production. Summary of the Invention

[0007] This invention aims to address the problems of low efficiency and easy loss of information in manual process reporting during precast slab production, as well as the high cost of RFID technology and the inability of traditional tracking methods to track precast slabs across regions. By deploying process cameras to collect images, training a model to identify process targets, and combining this with edge computing units to achieve edge deployment of the model, a cross-camera trajectory association method is constructed to realize automatic process information reporting and cross-regional tracking of precast slabs. This drives a digital twin model to display production status in real time, improving the automation and intelligence level of precast slab production management.

[0008] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides an intelligent detection method for precast beam and slab processes based on target detection, comprising:

[0009] S1. Deploy process cameras according to the precast slab production line and define the world coordinate system to complete the calibration of the deployed process cameras;

[0010] S2. Train the recognition model to identify the process target, and form a short-term trajectory by linking the detection results of consecutive frames of the same camera through the tracking method. Record the spatiotemporal motion characteristics of the target and convert the pixel coordinates into world coordinates through camera inverse projection.

[0011] S3. When a target leaves the overlapping area of ​​camera A and camera B, start a matching timer; and set a window time threshold. Determine the correlation between the trajectory of camera A and the new trajectory of camera B by searching for the newly emerging trajectory in the overlapping area of ​​camera B and checking whether its initial position is within the set range of the target prediction of camera A and comparing the motion characteristics of the two; and assign the same global ID to the associated trajectories.

[0012] S4. Complete the hardware construction of the intelligent data entry system for precast slab processes and determine the correspondence between process targets and precast slab processes; when "reinforcing cage hoisting" is detected, a virtual ID is generated, and other processes retrieve the virtual ID based on time, complete the process information entry and encrypt the primary key;

[0013] S5. Drive the digital twin model of the precast concrete plant through the test results to display the production status in real time.

[0014] Furthermore, the specific definition method of the world coordinate system in S1 is as follows:

[0015] The camera closest to the midpoint of the overhead plane is selected as the calibration reference camera, and the ground projection of the optical center of the calibration reference camera is taken as the origin of the world coordinate system.

[0016] Furthermore, the specific method for calibrating the process cameras deployed in S1 is as follows:

[0017] Let the set of cameras involved in the calibration of the deployed process cameras be T = {C1, C2, ..., C...} n}, where n represents the total number of cameras deployed, and C is the selected calibration reference camera. O C O ∈T;

[0018] For each camera C in set T i The set of cameras in the overlapping region can be represented as:

[0019] S i ={C j |C j ∈T,j≠i,C i With C j The monitored areas overlap (i = 1, 2, ..., n).

[0020] Using the calibration reference camera C O Starting from the grid, C is calculated using the checkerboard calibration method. O The set S contains the internal parameters K and extrinsic parameters of all cameras, and the accuracy of the parameters is optimized by averaging the values ​​through multiple registrations; where the set S... O Indicates the reference camera S for calibration O The monitored area consists of a collection of all other cameras that overlap with it;

[0021] Construct a set of calibrated cameras C calibrated Initially, C O With S O Cameras that have completed calibration are included in this set, that is:

[0022] C calibrated ={C O}∪{C j |C j ∈S O Internal parameter calibration has been completed.

[0023] Filter the set of unlabeled cameras C uncalibrated Defined as:

[0024]

[0025] From set C uncalibratedChoose one camera C i As a new calibration reference camera;

[0026] Construct a new set of cameras to be calibrated:

[0027]

[0028] That is, to filter out those related to C i There are overlapping and unlabeled cameras;

[0029] For S i The cameras in ' are respectively connected to C i Perform chessboard calibration, calculate internal parameters and take the mean, and then apply C... calibrated And C uncalibrated To update, that is:

[0030] C calibrated =C calibrated ∪{C i}∪S i '

[0031]

[0032] Repeat the above steps until... All cameras have been calibrated.

[0033] Furthermore, the method for calculating the extrinsic parameters is as follows:

[0034] Using camera A as the calibration reference camera, the translation vector T of camera A about the origin of the world coordinate system is... A =[X A ,Y A H A ] T , where X A ,Y A The X-axis and Y-axis coordinates in the world coordinate system, H A The height of camera A;

[0035] Place a chessboard calibration board in the field of view of camera A. The chessboard is placed on the ground, so its Z=0. Obtain the world coordinates P of the corner points of the chessboard. w Image coordinates p of camera A A Solving the rotation matrix R using the PnP method A ,satisfy:

[0036] s·p A =K A ·[R A |T A ]·P w

[0037] KA s is the intrinsic parameter of camera A, and s is the scale factor;

[0038] Next, place the calibration board in the overlapping area of ​​camera A and camera B, which needs to be calibrated as a reference, and simultaneously acquire images from both cameras. Then, solve the rotation matrix R of camera B relative to A using dual-camera calibration. BA Translation matrix T BA ,satisfy:

[0039] s·p B =K B ·[R BA |T BA ]·P cA

[0040] p B Let P be the image coordinates of B. cA Let A be the coordinates of the point in coordinate system A.

[0041] From the coordinate transformation relationship, we get:

[0042] R B =R BA ·R A

[0043] T B =R BA ·T A +T BA

[0044] Among them, R B and T B Let be the rotation matrix and translation vector of camera B in the world coordinate system.

[0045] Furthermore, the specific method for converting pixel coordinates to world coordinates through camera inverse projection in step S2 is as follows:

[0046] The process target is located on the ground, Z w Setting the value to 0 simplifies the 3D point projection problem into a system of 2D linear equations. The camera projection equations are as follows:

[0047]

[0048] Where: (u,v) are pixel coordinates, K is the camera intrinsic parameter matrix, R is the rotation matrix, T is the translation vector, (X... w ,Y w Z w ) is the world coordinate, s is the depth factor; f x f y c represents the focal length of the camera along the x-axis and y-axis, respectively; x c yThe coordinates of the principal point of the image are the coordinates of the intersection of the camera's optical axis and the image plane in the pixel coordinate system, representing the center position of the image; r ij (i = 1, 2, 3; j = 1, 2) represent the elements in the rotation matrix R; t x t y t z Let t be the components of the translation vector T, describing the translation relationship between the origin of the world coordinate system and the origin of the camera coordinate system. x t y t z These represent the translation distances along the x, y, and z axes, respectively; X W ,Y W Here, Z represents the horizontal coordinates of the target point in the world coordinate system. Since the target is located on the ground, Z is... W =0;

[0049] The results were:

[0050] s·u=f x (r 11 X W +r 12 Y W +t x )+c x (r 31 X W +r 32 Y W +t z )

[0051] s·v=f y (r 21 X W +r 22 Y W +t y )+c y (r 31 X W +r 32 Y W +t z )

[0052] s = r 31 X W +r 32 Y W +t z

[0053] Substituting s into the two equations and simplifying, we get:

[0054]

[0055] Here, A1 and A2 are the elements of the first column of the coefficient matrix, obtained by combining the camera's intrinsic and extrinsic parameters with pixel coordinates, reflecting the world coordinates X. W The coefficients affecting the projection relationship of pixel coordinates; B1, B2, are elements in the second column of the coefficient matrix, also obtained by calculation from camera intrinsic and extrinsic parameters and pixel coordinates, reflecting the world coordinates Y. W The influence coefficients on the pixel coordinate projection relationship; C1 and C2 are elements of the vector on the right side of the equation, also derived from camera intrinsic and extrinsic parameters and pixel coordinates through calculations. They are calculation results related to pixel coordinates, representing the values ​​determined based on known pixel coordinates and camera parameters under the current projection relationship, used to construct equations to solve for X. W and Y W The details are as follows:

[0056] A1=ur 31 -f x r 11 -c x r 31

[0057] B1 = ur 32 -f x r 12 -c x r 32

[0058] C1 = f x t x +c x t z -ut z

[0059] A2=ur 31 -f y r 21 -c y r 31

[0060] B2 = ur 32 -f y r 22 -c y r 32

[0061] C2 = f y t y +c y t z -ut z

[0062] After sorting, we can obtain:

[0063]

[0064]

[0065] That is, the two-dimensional coordinates of the target in the world coordinate system can be obtained by solving the problem, thus avoiding the difficulty of directly estimating the depth s.

[0066] Furthermore, the specific method for determining the correlation between the trajectory of camera A and the new trajectory of camera B in step S3 is as follows:

[0067] When the target leaves the overlapping area of ​​camera A: record the departure time t. leave_A , leaving position (X) leave_A ,Y leave_A ), and motion characteristics: velocity v leave_A Direction θ leave_A And start the matching timer T match ;

[0068] Based on the target's motion state when it leaves camera A, predict its possible location in the overlapping area of ​​camera B:

[0069] X pred_B =X leave_A +v leave_A ·cos(θ leave_A )·Δt

[0070]

[0071] Where, Δt=tt leave_A The time interval after leaving;

[0072] The matching window is set as follows:

[0073] Time window: [t leave_A ,t leave_A +T threshold ];

[0074] Spatial window: a region R centered on the predicted location. pred_B ;

[0075] Feature window: velocity deviation threshold Δv threshold directional deviation threshold Δθ threshold ;

[0076] Within the time window, examine each newly emerging trajectory T in the overlapping area of ​​camera B. new_B Perform the following matching:

[0077] Initial position matching: Check T new_B Is the initial position in R? pred_B Inside;

[0078] Time consistency: Check T new_B Does the time of appearance fall within [t]?leave_A ,t leave_A +T threshold ]Inside;

[0079] Motion feature matching: Calculating initial velocity and direction Check if the following conditions are met:

[0080]

[0081]

[0082] For each candidate trajectory T new_B Calculate the matching score S; and select the trajectory with the highest score. Match its score S with the set S threshold Compare, if S≥S threshold If so, the association is confirmed, and the same global ID is assigned.

[0083] Furthermore, the matching score S is calculated as follows:

[0084] The matching score S is a weighted scoring function that integrates multi-dimensional feature similarity and is used to quantify the likelihood of association between two trajectories. The following are the specific calculation formulas and implementation details for each dimension's score:

[0085] Location similarity S position The spatial proximity between the predicted and actual detected locations is measured using a normalized form of Gaussian kernel function / Euclidean distance:

[0086]

[0087] in:

[0088]

[0089] d is the Euclidean distance between the predicted position and the detected position, X detect_B ,Y detect_B σ represents the x and y coordinates of the detection location, respectively; σ is the standard deviation of the location.

[0090] Time Consistency S time Assess whether the time interval between the target leaving camera A and appearing at camera B matches the expected motion:

[0091]

[0092] Where, Δt=t detect_B -t leave_A This is the actual time interval; The theoretical time predicted based on velocity; d ABτ represents the spatial distance from camera A to camera B obtained through calibration; τ is the set time tolerance.

[0093] Speed ​​similarity S velocity Taking into account the consistency of both the magnitude and direction of the velocity:

[0094] S velocity =w v ·S v +w θ ·S θ

[0095] Among them, the similarity in speed magnitude S v for:

[0096]

[0097] In the formula, σ v The standard deviation of the velocity;

[0098] Speed ​​Magnitude Similarity S v for:

[0099]

[0100] In the formula, σ θ The standard deviation of the velocity;

[0101] Weighting coefficient: w v +w θ =1;

[0102] The comprehensive matching score is calculated as follows:

[0103] S = w1·S position +w2·S time +w3·S velocit

[0104] Where w1, w2, and w3 are the weights of positional similarity, temporal consistency, and velocity similarity, respectively, and w1 + w2 + w3 = 1.

[0105] Furthermore, the hardware construction of the intelligent reporting system for the precast slab process in S4 is as follows:

[0106] The precast slab process intelligent reporting system includes process cameras, switches, hard disk recorders, and edge computing units. By installing process cameras in various production areas of the precast slab, construction images are collected in real time. The edge computing unit can read the monitoring images via a local area network through the switch. The edge computing unit performs inference calculations to identify the precast slab process and uploads the process information to the cloud server to complete the process reporting.

[0107] As a second aspect of the present invention, an intelligent inspection system for precast beam and slab processes based on target detection is also provided, comprising:

[0108] Define a world coordinate system and camera calibration unit to deploy process cameras according to the precast slab production line. The camera closest to the midpoint of the overhead plane is used as the calibration reference camera. The world coordinate system is defined with the ground projection of the optical center of the calibration reference camera as the origin, and the calibration of the deployed process cameras is completed.

[0109] The process target recognition, tracking and coordinate transformation unit is used to train the recognition model to recognize process targets. It forms a short-term trajectory by associating the detection results of consecutive frames from the same camera through the tracking method, records the spatiotemporal motion characteristics of the target, and converts the pixel coordinates into world coordinates through camera inverse projection.

[0110] The cross-camera trajectory association and ID allocation unit is used to start a matching timer when a target leaves the overlapping area of ​​camera A and camera B; and set a window time threshold. It determines the association between the trajectory of camera A and the new trajectory of camera B by searching for the newly emerging trajectory in the overlapping area of ​​camera B and checking whether its initial position is within the set range of the target prediction of camera A and comparing the motion characteristics of the two; and assigns the same global ID to the associated trajectory.

[0111] The process target association and information filling encryption unit is used to determine the correspondence between process targets and precast slab processes. When "reinforcing cage hoisting" is detected, a virtual ID is generated. Other processes retrieve the virtual ID based on time, complete the process information filling and encrypt the primary key.

[0112] The digital twin model display unit is used to drive the digital twin model of the precast concrete plant through the test results, and to display the production status in real time.

[0113] As a third aspect of the present invention, a computer-readable storage medium is also provided, on which a computer program is stored, wherein the computer program is executed by a processor of any step of the above-described intelligent detection method for precast beam slab processes based on target detection.

[0114] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0115] 1. The intelligent detection method for precast beam slab processes based on target detection of the present invention deploys process cameras on the precast slab production line, using the camera closest to the midpoint of the overhead view plane as a calibration reference. A world coordinate system is defined with its optical center ground projection as the origin, and camera calibration is completed, achieving unified calibration of the spatial coordinates of multiple cameras. This technical feature utilizes the principle of geometric calibration to ensure that the target position under the view of each camera can be mapped to the same world coordinate system, solving the problem of inconsistent target positioning across cameras and laying a spatial benchmark for subsequent trajectory association and process analysis.

[0116] 2. The intelligent detection method for precast beam slab processes based on target detection of the present invention identifies process targets by training a recognition model, and combines a tracking method to associate the detection results of consecutive frames from the same camera to form a short-term trajectory. Simultaneously, it records the spatiotemporal motion characteristics of the target and converts pixel coordinates into world coordinates through camera inverse projection, achieving dynamic tracking and spatial positioning of the process target. This technical feature utilizes target detection and computer vision methods to achieve the identification and trajectory association of process targets. Simultaneously, it simplifies 3D projection calculations through ground constraints, enabling real-time acquisition of the target's position and motion status on the production line, providing data support for capturing process time nodes.

[0117] 3. The intelligent detection method for precast beam slab production based on target detection of the present invention starts a matching timer when the target leaves the overlapping area of ​​the cameras, sets a time threshold and searches for new trajectories, and achieves cross-camera trajectory association based on position prediction range and motion feature comparison, assigning global IDs to associated trajectories; simultaneously, it constructs an intelligent data entry system hardware, generates virtual IDs based on specific process detection, and completes the encryption of other process information through time retrieval, ultimately driving a digital twin model to display the production status in real time. This technical feature achieves cross-camera trajectory association through a multi-dimensional matching strategy, combines encryption technology to ensure process data security, and achieves visualized presentation of production progress through digital twin-driven methods, improving the intelligent management level of precast beam slab production. Attached Figure Description

[0118] Figure 1 This is a flowchart of the intelligent detection method for precast beam and slab processes based on target detection, according to an embodiment of the present invention.

[0119] Figure 2 This is a schematic diagram of process target recognition and single-camera tracking according to an embodiment of the present invention;

[0120] Figure 3 This is a flowchart of the intelligent data entry technology for prefabricated bridge deck processes according to an embodiment of the present invention;

[0121] Figure 4 This is a schematic diagram of the installation of a construction area process camera according to an embodiment of the present invention;

[0122] Figure 5 This is a schematic diagram illustrating the division of a precast slab production line according to an embodiment of the present invention;

[0123] Figure 6 This is a schematic diagram of the steel mesh dataset according to an embodiment of the present invention;

[0124] Figure 7 This is a schematic diagram of the model training process according to an embodiment of the present invention;

[0125] Figure 8 This is a schematic diagram of the hardware structure of an embodiment of the present invention;

[0126] Figure 9 This is a schematic diagram illustrating the determination of the target area for a process according to an embodiment of the present invention;

[0127] Figure 10 This is a schematic diagram of the precast panel process identification results and code binding in an embodiment of the present invention;

[0128] Figure 11 This is a schematic diagram of a board factory twin model according to an embodiment of the present invention;

[0129] Figure 12 This is a system unit diagram of an embodiment of the present invention. Detailed Implementation

[0130] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0131] Example 2

[0132] Please refer to Figure 1 This embodiment 1 provides an intelligent detection method for precast beam and slab processes based on target detection, including:

[0133] S1. Deploy process cameras according to the precast slab production line, take the camera closest to the midpoint of the overhead plane as the calibration reference camera, define the world coordinate system with the ground projection of the optical center of the calibration reference camera as the origin, and complete the calibration of the deployed process cameras;

[0134] S2. Train the recognition model to identify the process target, and form a short-term trajectory by linking the detection results of consecutive frames of the same camera through the tracking method. Record the spatiotemporal motion characteristics of the target and convert the pixel coordinates into world coordinates through camera inverse projection.

[0135] S3. When a target leaves the overlapping area of ​​camera A and camera B, start a matching timer; and set a window time threshold. Determine the correlation between the trajectory of camera A and the new trajectory of camera B by searching for the newly emerging trajectory in the overlapping area of ​​camera B and checking whether its initial position is within the set range of the target prediction of camera A and comparing the motion characteristics of the two; and assign the same global ID to the associated trajectories.

[0136] S4. Complete the hardware construction of the intelligent data entry system for precast slab processes and determine the correspondence between process targets and precast slab processes; when "reinforcing cage hoisting" is detected, a virtual ID is generated, and other processes retrieve the virtual ID based on time, complete the process information entry and encrypt the primary key;

[0137] S5. Drive the digital twin model of the precast concrete plant through the test results to display the production status in real time.

[0138] The above steps will be further explained in the following embodiment 1.

[0139] (1) Define the world coordinate system and camera calibration

[0140] The camera closest to the midpoint of the overhead plane is selected as the calibration reference camera. The ground projection of the optical center of the calibration reference camera is taken as the origin of the world coordinate system. In a specific preferred embodiment, the direction of movement of the precast slab in the slab plant is selected as the positive X-axis, the direction perpendicular to the X-axis toward the curing chamber is selected as the Y-axis, and the Z-axis is perpendicular to the ground and upward.

[0141] In a preferred embodiment, the specific method for calibrating the deployed process cameras is as follows:

[0142] Let the set of cameras involved in the calibration of the deployed process cameras be T = {C1, C2, ..., C...} n}, where n represents the total number of cameras deployed, and C is the selected calibration reference camera. O C O ∈T;

[0143] For each camera C in set T i The set of cameras in the overlapping region can be represented as:

[0144] S i ={C j |C j ∈T,j≠i,C i With C j The monitored areas overlap (i = 1, 2, ..., n).

[0145] Using the calibration reference camera C O Starting from the grid, C is calculated using the checkerboard calibration method.O The set S contains the internal parameters K and extrinsic parameters of all cameras, and the accuracy of the parameters is optimized by averaging the values ​​through multiple registrations; where S is the set. O Indicates the reference camera S for calibration O The monitored area consists of a collection of all other cameras that overlap with it;

[0146] Construct a set of calibrated cameras C calibrated Initially, C O With S O Cameras that have completed calibration are included in this set, that is:

[0147] C calibrated ={C O}∪{C j |C j ∈S o Internal parameter calibration has been completed.

[0148] Filter the set of unlabeled cameras C uncalibrated Defined as:

[0149]

[0150] From set C uncalibrated Choose one camera C i As a new calibration reference camera;

[0151] Construct a new set of cameras to be calibrated:

[0152]

[0153] That is, to filter out those related to C i There are overlapping and unlabeled cameras;

[0154] For S i The cameras in ' are respectively connected to C i Perform chessboard calibration, calculate internal parameters and take the mean, and then apply C... calibrated And C uncalibrated To update, that is:

[0155] C calibrated =C calibrated ∪{C i}∪S i '

[0156]

[0157] Repeat the above steps until... All cameras have been calibrated.

[0158] In a preferred embodiment, the external parameters are calculated as follows:

[0159] Using camera A as the calibration reference camera, the translation vector T of camera A about the origin of the world coordinate system is... A =[X A ,Y A H A ] T , where X A ,Y A The X-axis and Y-axis coordinates in the world coordinate system, H A The height of camera A;

[0160] Place a chessboard calibration board in the field of view of camera A. The chessboard is placed on the ground, so its Z=0. Obtain the world coordinates P of the corner points of the chessboard. w Image coordinates p of camera A A Solving the rotation matrix R using the PnP method A ,satisfy:

[0161] s·p A =K A ·[R A |T A ]·P w

[0162] K A s is the intrinsic parameter of camera A, and s is the scale factor;

[0163] Next, place the calibration board in the overlapping area of ​​camera A and camera B, which needs to be calibrated as a reference, and simultaneously acquire images from both cameras. Then, solve the rotation matrix R of camera B relative to A using dual-camera calibration. BA Translation matrix T BA ,satisfy:

[0164] s·p B =K B ·[R BA |T BA ]·P cA

[0165] p B Let P be the image coordinates of B. cA Let A be the coordinates of the point in coordinate system A.

[0166] From the coordinate transformation relationship, we get:

[0167] R B =R BA ·R A

[0168] T B =R BA ·T A +TBA

[0169] Among them, R B and T B Let be the rotation matrix and translation vector of camera B in the world coordinate system.

[0170] (2) Process target identification, tracking and coordinate transformation

[0171] Train the YOLO-V8 process target detection method to achieve process target recognition; please refer to... Figure 2 By using tracking methods (such as SORT, DeepSORT, ByteTrack, etc.), the detection results of process targets in consecutive frames from the same camera are correlated to form short-term target trajectories, and the time, position, speed, and direction of the target entering / leaving the overlapping area are recorded in real time. The trajectory points are converted from pixel coordinates to world coordinates through camera backprojection.

[0172] In a preferred embodiment, the specific method for converting pixel coordinates to world coordinates through camera inverse projection is as follows:

[0173] The process target is located on the ground, Z w Setting the value to 0 simplifies the 3D point projection problem into a system of 2D linear equations. The camera projection equations are as follows:

[0174]

[0175] Where: (u,v) are pixel coordinates, K is the camera intrinsic parameter matrix, R is the rotation matrix, T is the translation vector, (X... w ,Y w Z w ) is the world coordinate, s is the depth factor; f x f y c represents the focal length of the camera along the x-axis and y-axis, respectively; x c y The coordinates of the principal point of the image are the coordinates of the intersection of the camera's optical axis and the image plane in the pixel coordinate system, representing the center position of the image; r ij (i = 1, 2, 3; j = 1, 2) represent the elements in the rotation matrix R; t x t y t z Let t be the components of the translation vector T, describing the translation relationship between the origin of the world coordinate system and the origin of the camera coordinate system. x t y t z These represent the translation distances along the x, y, and z axes, respectively; X W ,Y W Here, Z represents the horizontal coordinates of the target point in the world coordinate system. Since the target is located on the ground, Z is...W =0;

[0176] The results were:

[0177] s·u=f x (r 11 X W +r 12 Y W +t x )+c x (r 31 X W +r 32 Y W +t z )

[0178] s·v=f y (r 21 X W +r 22 Y W +t y )+c y (r 31 X W +r 32 Y W +t z )

[0179] s = r 31 X W +r 32 Y W +t z

[0180] Substituting s into the two equations and simplifying, we get:

[0181]

[0182] Here, A1 and A2 are the elements of the first column of the coefficient matrix, obtained by combining the camera's intrinsic and extrinsic parameters with pixel coordinates, reflecting the world coordinates X. W The coefficients affecting the projection relationship of pixel coordinates; B1, B2, are elements in the second column of the coefficient matrix, also obtained by calculation from camera intrinsic and extrinsic parameters and pixel coordinates, reflecting the world coordinates Y. W The influence coefficients on the pixel coordinate projection relationship; C1 and C2 are elements of the vector on the right side of the equation, also derived from camera intrinsic and extrinsic parameters and pixel coordinates through calculations. They are calculation results related to pixel coordinates, representing the values ​​determined based on known pixel coordinates and camera parameters under the current projection relationship, used to construct equations to solve for X. W and Y W The details are as follows:

[0183] A1=ur 31 -fx r 11 -c x r 31

[0184] B1 = ur 32 -f x r 12 -c x r 32

[0185] C1 = f x t x +c x t z -ut z

[0186] A2=ur 31 -f y r 21 -c y r 31

[0187] B2 = ur 32 -f y r 22 -c y r 32

[0188] C2 = f y t y +c y t z -ut z

[0189] After sorting, we can obtain:

[0190]

[0191] That is, the two-dimensional coordinates of the target in the world coordinate system can be obtained by solving the problem, thus avoiding the difficulty of directly estimating the depth s.

[0192] (3) Cross-camera trajectory association and ID allocation

[0193] When a target leaves the overlapping area of ​​cameras A and B, a matching timer is started. Within a very short time window (the time threshold can be set based on the size of the overlapping area and the trolley's speed), a new trajectory is searched for in the overlapping area of ​​camera B. For the new trajectory, its initial position is checked to see if it is within a reasonable range of the target prediction from camera A (a distance threshold is set based on the projection transformation error). The motion characteristics (speed, direction) of the two are compared. If both are satisfied, the trajectory from camera A is associated with the new trajectory from camera B, and they are assigned the same global ID, thus enabling cross-camera target tracking.

[0194] In a preferred embodiment, the specific method for determining the correlation between the trajectory of camera A and the new trajectory of camera B is as follows:

[0195] When the target leaves the overlapping area of ​​camera A: record the departure time t. leave_A , leaving position (X) leave_A ,Y leave_A ), and motion characteristics: velocity v leave_A Direction θ leave_A And start the matching timer T match ;

[0196] Based on the target's motion state when it leaves camera A, predict its possible location in the overlapping area of ​​camera B:

[0197] X pred_B =X leave_A +v leave_A ·cos(θ leave_A )·Δt

[0198]

[0199] Where, Δt=tt leave_A The time interval after leaving;

[0200] The matching window is set as follows:

[0201] Time window: [t leave_A ,t leave_A +T threshold ];

[0202] Spatial window: a region R centered on the predicted location. pred_B ;

[0203] Feature window: velocity deviation threshold Δv threshold directional deviation threshold Δθ threshold ;

[0204] Within the time window, examine each newly emerging trajectory T in the overlapping area of ​​camera B. new_B Perform the following matching:

[0205] Initial position matching: Check T new_B Is the initial position in R? pred_B Inside;

[0206] Time consistency: Check T new_B Does the time of appearance fall within [t]? leave_A ,t leave_A +T threshold ]Inside;

[0207] Motion feature matching: Calculating the initial velocity v newBand direction Check if the following conditions are met:

[0208]

[0209]

[0210] For each candidate trajectory T new_B Calculate the matching score S; and select the trajectory with the highest score. Match its score S with the set S threshold Compare, if S≥S threshold If so, the association is confirmed, and the same global ID is assigned.

[0211] In a preferred embodiment, the matching score S is calculated as follows:

[0212] The matching score S is a weighted scoring function that integrates multi-dimensional feature similarity and is used to quantify the likelihood of association between two trajectories. The following are the specific calculation formulas and implementation details for each dimension's score:

[0213] Location similarity S position The spatial proximity between the predicted and actual detected locations is measured using a normalized form of Gaussian kernel function / Euclidean distance:

[0214]

[0215] in:

[0216]

[0217] d is the Euclidean distance between the predicted position and the detected position, X detect_B ,Y detect_B σ represents the x and y coordinates of the detection location, respectively; σ is the standard deviation of the location.

[0218] Time Consistency S time Assess whether the time interval between the target leaving camera A and appearing at camera B matches the expected motion:

[0219]

[0220] Where, Δt=t detect_B -t leave_A This is the actual time interval; The theoretical time predicted based on velocity; d AB τ represents the spatial distance from camera A to camera B obtained through calibration; τ is the set time tolerance.

[0221] Speed ​​similarity S velocity Taking into account the consistency of both the magnitude and direction of the velocity:

[0222] S velocity =w v ·S v +w θ ·S θ

[0223] Among them, the similarity in speed magnitude S v for:

[0224]

[0225] In the formula, σ v The standard deviation of the velocity;

[0226] Speed ​​Magnitude Similarity S v for:

[0227]

[0228] In the formula, σ θ The standard deviation of the velocity;

[0229] Weighting coefficient: w v +w θ =1;

[0230] The comprehensive matching score is calculated as follows:

[0231] S = w1·S position +w2·S time +w3·S velocity

[0232] Where w1, w2, and w3 are the weights of positional similarity, temporal consistency, and velocity similarity, respectively, and w1 + w2 + w3 = 1.

[0233] (4) Linking process targets and encrypting information entry

[0234] 4.1 Camera Deployment in the Process

[0235] Please refer to Figure 3 Intelligent data entry for precast bridge deck construction processes requires the installation of process cameras in the construction area. Precast bridge decks are produced in a precast panel factory, and their main processes are as follows: reinforcement installation, material placement, vibration, steam curing, demolding, and storage. In a preferred embodiment, please refer to... Figure 4 Referring to the precast concrete construction plan and site drawings for the Shiziyang Channel, the installation locations and number of cameras for each process were determined. The cameras were installed at a height that ensured most of the trolley or precast slab was within the frame, guaranteeing a clear view of the rebar cage lifting equipment and the concrete placing boom.

[0236] In a specific preferred embodiment, the specific hardware devices and parameters of the camera are shown in Table 1:

[0237] Table 1. List of Camera Equipment Installed in Each Process

[0238]

[0239] 4.2 Construction of Precast Slab Process Dataset

[0240] Please refer to Figure 5 In a specific preferred embodiment, based on the construction plan and construction site drawings of the precast concrete slabs for the Shiziyang Channel, the division of each process in the precast concrete slab production line was drawn, and the ID of each workstation was set.

[0241] By observing the precast panel production process, in a specific preferred embodiment, the possible process objectives for each process were determined, as shown in Table 2.

[0242] Table 2 Construction Targets for Precast Slabs

[0243]

[0244] The processes of precast slab reinforcement construction, material placement, vibration, steam curing, and demolding can be identified through six types of process targets: reinforcement cage hoisting, reinforcement cage installation, concrete pouring, precast slab installation, precast slab loading and unloading, and trolley installation. These six process targets, including reinforcement cage hoisting and trolley installation, are used as process identification targets for precast bridge decks, and image data of the precast bridge deck production stages are collected.

[0245] To ensure higher dataset quality and closer resemblance to actual precast panel production process images, image data was captured using process cameras on the precast bridge panel production line. This ensured the images covered all stages of the bridge panel production process, and image acquisition was conducted under varying lighting and background conditions. Please refer to [reference needed]. Figure 6 After data collection is completed, the image data is cleaned. In a specific preferred embodiment, more than 5,000 images are obtained, including images of the above six types of process targets.

[0246] 4.3 YOLOv8 Model

[0247] The automatic reporting of process steps requires real-time detection of process information. When selecting a target detection network, both recognition accuracy and inference speed must be considered. In a preferred embodiment, real-time target detection is achieved using YOLOv8; YOLOv8 includes an input layer, a backbone layer, a feature fusion neck layer, and a prediction output head layer.

[0248] YOLOv8 has improved and optimized several aspects compared to previous generations of the YOLO series. In the backbone network, YOLOv8 retains the CSP architecture framework of YOLOv5, but replaces the C3 module with the C2f module, further reducing weight while improving feature extraction capabilities. At the prediction output end, an Anchor-Free + Decoupled-head design is introduced to adapt to objects of different scales and shapes. In the loss function, VFLLoss is used for classification, and DFLLoss + CIoULoss are used for regression.

[0249] 4.4 Model Training Results

[0250] In a preferred embodiment, the process identification dataset is divided into training, validation, and test sets in a 7:1:2 ratio, and iterative training is performed on the dataset. The model is trained on an Ubuntu 22.04 system and an NVIDIA RTX 3090 GPU. After multiple training iterations, two weight files are obtained: last.pt contains the latest weights, and best.pt contains the optimal weights. The training results are as follows: Figure 7 As shown, the network tends to converge after about 80 iterations.

[0251] 4.5 System Architecture Design

[0252] Please refer to Figure 8 The entire precast slab process intelligent reporting system consists of process cameras, switches, hard disk recorders, and edge computing units (NVIDIA Jetson Orin Nano). Process cameras are installed in various precast slab production areas to capture construction images in real time. The edge computing unit can read the monitoring images via a local area network through the switches. The edge computing unit performs inference calculations to identify the precast slab process and uploads the process information to the cloud server, completing the process reporting.

[0253] In a preferred embodiment, the NVIDIA Jetson Orin Nano was selected as the edge computing unit. Jetson Nano is an embedded development platform based on NVIDIA GPUs, primarily designed for deep learning, computer vision, and embedded application development. Jetson Nano's computing power and processing speed make it one of the preferred platforms in the fields of deep learning and computer vision. The Jetson Nano product parameters used in a preferred embodiment are shown in Table 3 below.

[0254] Table 3 Jetson Nano Product Specifications

[0255]

[0256] 4.6 Accelerated Model Inference

[0257] In a preferred embodiment, since automatic process reporting requires real-time identification of process targets, the optimized process target identification model needs to be quantized using TensorRT to minimize model inference time while maintaining acceptable accuracy. TensorRT is a high-performance deep learning deployment and inference optimization library provided by NVIDIA. It improves model inference efficiency through techniques such as kernel fusion, weight quantization, and graph optimization. TensorRT can be used in conjunction with mainstream deep learning frameworks such as PyTorch and supports ONNX format models for easy model conversion and optimization. The following commands are used to quantize the model in JetsonNano to fp16 accuracy.

[0258] 4.7 Process Information Reporting

[0259] In a specific preferred embodiment, the correspondence between process targets and precast panel processes in each work area is shown in Table 4 below. When a specific process target is detected, it is determined that the panel is in the corresponding process.

[0260] Table 4. Correspondence between process objectives and precast slab processes

[0261]

[0262] Please refer to Figure 9 When the same process target corresponds to multiple precast slab processes, such as precast slab target corresponding to multiple processes such as material laying and vibration, the process target - workstation ID is used as the process target. For example, precast slab-7 represents the end of the material laying process, and precast slab-8 and precast slab-9 represent the start and end of the vibration process, respectively. This article divides the process camera image into different zones to determine the work area where the precast slab target is located, as shown in the figure below.

[0263] Please refer to Figure 10In a preferred embodiment, the rebar tying process is set as the initial process. When the "rebar cage hoisting" target is detected, it signifies the start of production for a slab, with the production date minus the production time serving as the unique virtual ID for that slab. The primary key, slab virtual ID, process ID, platform, process time, and identification image are uploaded to the cloud database as process information to complete the process reporting. When other process targets are detected, the previous process record is retrieved based on the process time to obtain the slab virtual ID, and the current process is designated as the subsequent process for that slab to complete the process reporting. Due to the complex on-site environment, factors such as lighting and equipment may interfere with the identification of process targets. In practical applications, the beam slab virtual ID and process ID are encrypted using MD5 and used as the primary key for the process record. Each time a process is reported, a search is performed to ensure there are no duplicate primary keys before inserting the record. The construction unit can bind the actual structure code via a WeChat mini-program, reattaching the process information linked to the virtual slab number to the actual structure to complete the process information reporting.

[0264] (5) Digital twin model display

[0265] Please refer to Figure 11 By leveraging the detection results of the YOLOv8 process identification model, the digital twin model of the precast concrete plant can be driven, displaying the on-site production situation in real time. During model inference, after identifying the process targets, each process target is mapped to the corresponding workstation, forming a process set for the entire precast concrete plant. This process set is then pushed to the precast concrete plant's smart screen backend via an interface, driving the twin model. Construction managers can intuitively grasp the on-site construction situation through the plant's twin model.

[0266] Based on the above, this embodiment 1 proposes an intelligent detection technology for precast slab processes based on target detection methods. This technology identifies process targets in real time through target detection, enabling automatic reporting of precast slab process information. First, based on the deployment of process recognition cameras at the precast slab plant site, the process targets corresponding to each process were determined, and a process recognition dataset was collected and constructed. Then, the YOLOv8 target detection method was trained to identify process targets, JetsonNano was selected as the edge computing unit, and the process recognition model was deployed. The computing unit was connected to the upstream switch of the video recorder via a local area network to identify processes in real time and push process information. Based on the process information, the plant's twin model is displayed in real time. Simultaneously, the intelligent reporting system for precast slab processes has been applied at the Shiziyang Channel precast slab plant, verifying the system's effectiveness.

[0267] Example 2

[0268] Please refer to Figure 12 This embodiment 2 provides an intelligent inspection system for precast beam and slab processes based on target detection, including:

[0269] Define a world coordinate system and camera calibration unit to deploy process cameras according to the precast panel production line and define the world coordinate system to complete the calibration of the deployed process cameras;

[0270] The process target recognition, tracking and coordinate transformation unit is used to train the recognition model to recognize process targets. It forms a short-term trajectory by associating the detection results of consecutive frames from the same camera through the tracking method, records the spatiotemporal motion characteristics of the target, and converts the pixel coordinates into world coordinates through camera inverse projection.

[0271] The cross-camera trajectory association and ID allocation unit is used to start a matching timer when a target leaves the overlapping area of ​​camera A and camera B; and set a window time threshold. It determines the association between the trajectory of camera A and the new trajectory of camera B by searching for the newly emerging trajectory in the overlapping area of ​​camera B and checking whether its initial position is within the set range of the target prediction of camera A and comparing the motion characteristics of the two; and assigns the same global ID to the associated trajectory.

[0272] The process target association and information filling encryption unit is used to determine the correspondence between process targets and precast slab processes. When "reinforcing cage hoisting" is detected, a virtual ID is generated. Other processes retrieve the virtual ID based on time, complete the process information filling and encrypt the primary key.

[0273] The digital twin model display unit is used to drive the digital twin model of the precast concrete plant through the test results, and to display the production status in real time.

[0274] Example 3

[0275] This embodiment 3 also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it can implement any step of an intelligent detection method for precast beam and slab processes based on target detection.

[0276] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0277] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.

[0278] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart detection method for precast beam and slab processes based on target detection, characterized in that, include: S1. Deploy process cameras according to the precast slab production line and define the world coordinate system to complete the calibration of the deployed process cameras; S2. Train the recognition model to identify the process target, and form a short-term trajectory by linking the detection results of consecutive frames of the same camera through the tracking method. Record the spatiotemporal motion characteristics of the target and convert the pixel coordinates into world coordinates through camera inverse projection. S3. When a target leaves the overlapping area of ​​camera A and camera B, start a matching timer; A window time threshold is set, and the correlation between the trajectory of camera A and the new trajectory of camera B is determined by searching for newly emerging trajectories in the overlapping area of ​​camera B and checking whether their initial positions are within the set range of the target prediction of camera A, as well as comparing the motion characteristics of the two; and the same global ID is assigned to the associated trajectories. S4. Complete the hardware construction of the intelligent data entry system for precast slab processes and determine the correspondence between process targets and precast slab processes; when "reinforcing cage hoisting" is detected, a virtual ID is generated, and other processes retrieve the virtual ID based on time, complete the process information entry and encrypt the primary key; S5. Drive the digital twin model of the precast concrete plant through the test results to display the production status in real time; The specific method for determining the correlation between the trajectory of camera A and the new trajectory of camera B is as follows: When the target leaves the overlapping area of ​​camera A: record the departure time. , leave position and motion characteristics: speed ,direction And start the matching timer ; Based on the target's motion state when it leaves camera A, predict its possible location in the overlapping area of ​​camera B: wherein, is the time interval after departure; The matching window is set as follows: Time window: ; Spatial window: region centered on predicted position ; Feature window: speed deviation threshold , direction deviation threshold ; Within the time window, check each newly appearing track in the camera B overlap region , the following matching is performed: Initial position match: check if initial position is within ;​ Temporal consistency: check if the occurrence time of is within ; Motion feature matching: Compute initial velocity and direction Check if it is satisfied: For each candidate trajectory Calculate the matching score And select the trajectory with the highest score. Match it with a score With the setting Compare, if If so, the association is confirmed, and the same global ID is assigned.

2. The intelligent detection method for precast beam and slab processes based on target detection according to claim 1, characterized in that, The specific method for defining the world coordinate system in S1 is as follows: The camera closest to the midpoint of the overhead plane is selected as the calibration reference camera, and the ground projection of the optical center of the calibration reference camera is taken as the origin of the world coordinate system.

3. The intelligent detection method for precast beam and slab processes based on target detection according to claim 1, characterized in that, The specific method for calibrating the process cameras deployed in S1 is as follows: Let the set of cameras participating in the calibration be denoted as . ,in, This indicates the total number of cameras deployed, with the selected calibration reference camera being... , ; to the set where each camera The set of cameras whose overlapping regions can be represented as: Using a calibration reference camera Starting from the grid, the calculation is performed using the checkerboard calibration method. and sets Internal parameters of all cameras And extrinsic parameters, and optimize parameter accuracy by averaging multiple registrations; among them, the set Indicates and calibrates the reference camera The monitored area consists of a collection of all other cameras that overlap with it; Construct a set of calibrated cameras Initially and Cameras that have completed calibration are included in this set, that is: Filtering unlabeled camera sets Defined as: From the set Choose one camera As a new calibration reference camera; Construct a new set of cameras to be calibrated: i.e. to screen out those with cameras with overlap and no calibration; The cameras in respectively with calibration, calculate the internal parameters and take the average, and update and , that is: Repeat the above steps until , all cameras are calibrated.

4. The method of claim 3, wherein the method is characterized by: The method for calculating the external parameters is as follows: Using camera A as the calibration reference camera, the translation vector of camera A about the origin of the world coordinate system is... ,in, X-axis and Y-axis coordinates in the world coordinate system The height of camera A; A chessboard-patterned calibration board is placed in the field of view of camera A. The chessboard is placed on the ground. Then its... Obtain the world coordinates of the corner points of the chessboard. Image coordinates of camera A ,pass Method for solving rotation matrices ,satisfy: s is the intrinsic parameter of camera A, and s is the scale factor; Next, place the calibration board in the overlapping area of ​​camera A and camera B, which needs to be calibrated as a reference, and simultaneously acquire images from both cameras. Then, solve the rotation matrix of camera B relative to A using dual-calibration. Translation matrix ,satisfy: Let B be the image coordinates. Let A be the coordinates of the point in coordinate system A. From the coordinate transformation relationship, we get: in, and Let be the rotation matrix and translation vector of camera B in the world coordinate system.

5. The method of claim 3, wherein the method is characterized by: The specific method for converting pixel coordinates to world coordinates via camera inverse projection in step S2 is as follows: The process target is located on the ground. Setting the value to 0 simplifies the 3D point projection problem into a system of 2D linear equations. The camera projection equations are as follows: in: These are pixel coordinates. It is the camera intrinsic parameter matrix. It is a rotation matrix. It is a translation vector. It is a world coordinate system. It is a depth factor; , These are the focal lengths of the camera along the x-axis and y-axis, respectively; , These are the coordinates of the principal point of the image, that is, the coordinates of the intersection of the camera's optical axis and the image plane in the pixel coordinate system, representing the center position of the image; Represents the elements in the rotation matrix R; , Translation vector The component describes the translation relationship between the origin of the world coordinate system and the origin of the camera coordinate system. , These represent the translation distances along the x, y, and z axes, respectively. , Here are the horizontal coordinates of the target point in the world coordinate system. Since the target is located on the ground, here... ; The results were: Substituting s into the two equations and simplifying, we get: in, , , is the element in the first column of the coefficient matrix, obtained by combining camera intrinsic and extrinsic parameters and pixel coordinates through calculations, reflecting world coordinates. Influence coefficient on pixel coordinate projection relationship; , , is an element in the second column of the coefficient matrix, also obtained through calculations using camera intrinsic and extrinsic parameters and pixel coordinates, reflecting world coordinates. Influence coefficient on pixel coordinate projection relationship; , The elements of the vector on the right side of the equation are also derived from camera intrinsic and extrinsic parameters and pixel coordinates through calculations. They are calculation results related to pixel coordinates, representing values ​​determined based on known pixel coordinates and camera parameters under the current projection relationship, used to construct and solve the equation. and The details are as follows: After sorting, we can obtain: That is, the two-dimensional coordinates of the target in the world coordinate system can be obtained by solving the problem, thus avoiding the difficulty of directly estimating the depth s.

6. The method of claim 1, wherein the method is characterized by: The matching score The calculation method is: Match score It is a weighted scoring function that integrates multi-dimensional feature similarity to quantify the likelihood of association between two trajectories; the following are the specific calculation formulas and implementation details for the scores of each dimension: Position similarity The spatial closeness of the predicted position to the actual detected position is measured by a normalized version of the Gaussian kernel function / Euclidean distance: in: To predict the Euclidean distance between the location and the detection location, These are the x and y coordinates of the detection location, respectively; The standard deviation of the location; Temporal consistency Evaluate whether the time interval between the target leaving camera A and appearing at camera B is consistent with the motion expectations: in, This is the actual time interval; The theoretical time predicted based on the speed; τ represents the spatial distance from camera A to camera B obtained through calibration; τ is the set time tolerance. speed similarity Taking into account the consistency of both the magnitude and direction of the velocity: Among them, the similarity of speed magnitude for: In the formula, is the velocity standard deviation; Speed magnitude similarity is: In the formula, is the velocity standard deviation; Weighting coefficients: ; The comprehensive matching score is calculated as follows: in, , , These are the weights for positional similarity, temporal consistency, and velocity similarity, respectively. .

7. The method of claim 1, wherein the method is characterized by: The hardware construction details of the intelligent reporting system for the precast slab process in S4 are as follows: The precast slab process intelligent reporting system includes process cameras, switches, hard disk recorders, and edge computing units; by installing process cameras in various production areas of the precast slab, construction images are collected in real time; the edge computing units can read the monitoring images via a local area network through the switches; The edge computing unit performs inference calculations to identify the prefabrication process and uploads the process information to the cloud server to complete the process reporting.

8. A precast beam slab process intelligent detection system based on target detection, characterized in that, include: Define a world coordinate system and camera calibration unit to deploy process cameras according to the precast panel production line and define the world coordinate system to complete the calibration of the deployed process cameras; The process target recognition, tracking and coordinate transformation unit is used to train the recognition model to recognize process targets. It forms a short-term trajectory by associating the detection results of consecutive frames from the same camera through a tracking method, records the spatiotemporal motion characteristics of the target, and converts the pixel coordinates into world coordinates through camera inverse projection. A cross-camera trajectory association and ID allocation unit is used to start a matching timer when a target leaves the overlapping area of ​​camera A and camera B; A window time threshold is set, and the correlation between the trajectory of camera A and the new trajectory of camera B is determined by searching for newly emerging trajectories in the overlapping area of ​​camera B and checking whether their initial positions are within the set range of the target prediction of camera A and comparing the motion characteristics of the two; and the same global ID is assigned to the associated trajectories. The process target association and information filling encryption unit is used to determine the correspondence between process targets and precast slab processes. When "reinforcing cage hoisting" is detected, a virtual ID is generated. Other processes retrieve the virtual ID based on time, complete the process information filling and encrypt the primary key. The digital twin model display unit is used to drive the digital twin model of the precast panel factory through the test results, and to display the production status in real time; The specific method for determining the correlation between the trajectory of camera A and the new trajectory of camera B is as follows: When the target leaves the overlap area of camera A: record the leaving time , leaving position , and motion characteristics: speed , direction and start the matching timer ; Based on the target's motion state when it leaves camera A, predict its possible location in the overlapping area of ​​camera B: wherein, is the time interval after departure; The matching window is set as follows: Time window: ; Spatial window: region centered on predicted position ; Feature window: speed deviation threshold , direction deviation threshold ; Within the time window, examine each newly appearing trajectory in the overlapping area of ​​camera B. Perform the following matching: Initial position match: check if initial position is within ;​ Temporal consistency: check if the occurrence time of is within ; Motion feature matching: compute initial velocity and direction check if: For each candidate trajectory Calculate the matching score And select the trajectory with the highest score. Match it with a score With the setting Compare, if If so, the association is confirmed, and the same global ID is assigned.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor as described in any one of claims 1-7: an intelligent detection method for precast beam and slab processes based on target detection.

Images