Real-time detection system for construction quality based on computer vision

By constructing a computer vision-based real-time construction quality detection system, the problem of verifying the authenticity of video stream data in high-altitude operation environments was solved. This enabled the quantitative assessment of transmission pressure and illusion risks, ensuring data credibility and business continuity in harsh communication environments.

CN123068976APending Publication Date: 2026-11-10THE THIRD CONSTR OF CHINA CONSTR EIGHTH ENG BUREAU
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611100812.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-23
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In high-altitude construction scenarios, existing monitoring systems cannot effectively verify the authenticity of video stream data under harsh communication environments, and lack quantitative assessment of transmission environment pressure and generative reconstruction risks, resulting in a disconnect between the monitoring images and the actual construction status.

Method used

A real-time construction quality monitoring system based on computer vision is constructed. Through a data acquisition and management center, a transmission pressure assessment unit, an illusion risk analysis unit, a pattern decision unit, and a consistency verification unit, the system calculates the transmission pressure index and illusion index and determines the pattern selection. Combined with building information model, physical consistency verification is performed to ensure the authenticity and continuity of the data.

Benefits of technology

It achieves precise quantification of transmission environment pressure, suppresses the risk of visual illusions, ensures the credibility of monitoring data and business continuity in harsh communication environments, and can perform intelligent dual-mode switching when the network fluctuates, eliminate environmental interference and sensor failures, and ensure the physical authenticity of monitoring data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123068976A_ABST
    Figure CN123068976A_ABST
Patent Text Reader

Abstract

The present application relates to the technical field of intelligent construction and engineering quality detection, in particular to a construction quality real-time detection system based on computer vision; comprising a data acquisition management center, transmission pressure evaluation, illusion risk analysis, mode decision and consistency verification unit; the system obtains transmission pressure index and illusion index by analyzing original video stream and channel state; the core is to make game decision according to illusion index and transmission delay, to adaptively select video stream or topological feature stream transmission mode, and to perform physical consistency closed loop verification on heterogeneous data; when the structural topological consistency residual exceeds the threshold value, a fake fault signal is generated; the present application constructs a supply and demand coupling model, realizes accurate quantification of transmission pressure in a harsh communication environment, and guarantees the real-time performance and reliability of high-altitude operation detection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent construction and engineering quality inspection technology, specifically a real-time construction quality inspection system based on computer vision. Background Technology

[0002] In high-altitude construction scenarios, real-time quality monitoring heavily relies on the accurate transmission of video stream data. However, wireless communication links in such environments often exhibit large bandwidth fluctuations and unstable channel conditions, and are susceptible to multipath effects and environmental noise interference. To ensure transmission continuity, existing monitoring solutions typically employ high-compression coding or introduce generative models for image restoration and reconstruction when dealing with high transmission pressure. While this approach reduces bandwidth consumption, it can easily lead to visual illusions that do not conform to physical facts due to model inference biases when data is severely lacking, resulting in a disconnect between the monitored images and the actual construction status. Furthermore, traditional systems lack a quantitative assessment mechanism for transmission environment pressure and model generation risks, and cannot intelligently switch between video streams and topological feature streams based on channel conditions, making it difficult to effectively verify the physical consistency of data while ensuring real-time performance.

[0003] Therefore, how to construct a detection system that can sense transmission pressure in real time, quantify generative reconstruction risks, and realize dual-modal intelligent switching and physical authenticity verification in order to solve the problem of verifying the authenticity of monitoring data under harsh communication environments is an urgent technical problem to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a real-time construction quality inspection system based on computer vision. Specifically, the technical solution of this invention includes:

[0005] The data acquisition and management center is used to retrieve the original video stream data, channel status data, and transmission delay data from the high-altitude work surface, and send the original video stream data, channel status data, and transmission delay data to the transmission stress assessment unit.

[0006] The transmission stress assessment unit is used to perform transmission environment stress assessment and analysis on the raw video stream data and channel status data to obtain the transmission stress index.

[0007] The hallucination risk analysis unit is used to perform generative reconstruction risk assessment and feedback analysis based on the transmission pressure index to obtain the hallucination index;

[0008] The mode decision unit is used to perform game-theoretic transmission decision quantitative matching analysis on illusion index and transmission delay data to obtain mode selection discrimination value, and compare and analyze the mode selection discrimination value with the preset discrimination value threshold.

[0009] If the mode selection discrimination value is greater than or equal to the preset discrimination value threshold, a video stream transmission signal is generated;

[0010] If the mode selection discrimination value is less than the preset discrimination value threshold, a topology feature flow transmission signal is generated;

[0011] The consistency verification unit is used to perform physical consistency closed-loop verification analysis on the acquired heterogeneous data in response to the topological feature flow transmission signal, obtain the structural topological consistency residual, and compare and analyze the structural topological consistency residual with the preset yield strain threshold.

[0012] If the structural topology consistency residual is less than or equal to the preset yield strain threshold, a normal data signal is generated.

[0013] If the structural topology consistency residual is greater than the preset yield strain threshold, a fake fault signal is generated.

[0014] Preferably, the transmission environment stress assessment and analysis process is as follows:

[0015] Frequency domain transformation analysis is performed on the raw video stream data to extract normalized visual entropy, and dynamic gain coefficients are constructed based on the normalized visual entropy.

[0016] Obtain the baseline bit rate required for the minimum recognition accuracy of machine vision, and multiply the baseline bit rate with the dynamic gain coefficient to obtain the data requirement.

[0017] The channel state data is parsed to extract the instantaneous detection bandwidth of the physical link and the channel packet loss rate;

[0018] Obtain the preset protocol overhead efficiency factor, and multiply the instantaneous probe bandwidth, channel packet loss rate, and protocol overhead efficiency factor to obtain the effective carrying capacity;

[0019] Dividing the data demand by the effective carrying capacity yields the transmission pressure index.

[0020] Preferably, the generative reconstruction risk assessment feedback analysis process is as follows:

[0021] Obtain the preset model confidence decay constant, and obtain the duration between the arrival time of the most recent fully validated keyframe and the current time, and set the duration as the keyframe interval duration;

[0022] The model confidence decay constant, transmission pressure exponent, and key frame interval duration are multiplied together, and the exponential decay value is calculated based on the product result. The natural decay component is obtained by subtracting the exponential decay value from the value 1.

[0023] Get the currently received sparse feature vector, retrieve the historical feature vector library, calculate the cosine distance between the sparse feature vector and the historical feature vector library, and set the cosine distance as the feature vector cosine distance.

[0024] Obtain the preset feature drift weight coefficients, and multiply the feature drift weight coefficients with the cosine distance of the feature vector to obtain the semantic mutation component;

[0025] The illusion index is obtained by summing the natural decay component and the semantic mutation component.

[0026] Preferably, the game transmission decision quantification matching analysis process is as follows:

[0027] Obtain a preset hallucination index threshold, and subtract the hallucination index from the preset hallucination index threshold to obtain the reality difference value;

[0028] Obtain the preset authenticity utility weight, multiply the authenticity utility weight by the authenticity difference value, and obtain the authenticity utility value.

[0029] Obtain the maximum allowed latency limit for the task, divide the transmission latency data by the maximum latency limit, and obtain the normalized latency ratio.

[0030] Obtain the preset real-time penalty weight, and multiply the real-time penalty weight with the normalized latency ratio to obtain the real-time penalty value;

[0031] Subtracting the real-time penalty value from the real utility value yields the pattern selection discriminant value.

[0032] Preferably, the physical consistency closed-loop verification and analysis process is as follows:

[0033] Collect observable rigid connection edges within the current field of view, and obtain the real-time three-dimensional coordinate vectors of the first and second endpoints of the rigid connection edges.

[0034] Using the Euclidean distance operator, the distance between the real-time three-dimensional coordinate vectors of the first and second endpoints is calculated to obtain the real-time side length.

[0035] Retrieve the theoretical static length of the rigid connection edge in the preset building information model;

[0036] Subtract the theoretical static length from the real-time side length to obtain the length difference. Divide the length difference by the theoretical static length to obtain the unilateral strain deviation.

[0037] The unilateral strain deviation of all rigid connection edges within the field of view is averaged to obtain the structural topology consistency residual.

[0038] Preferably, the structural topology consistency residual is used to characterize the rate of change of the internal geometric invariants of the structural component;

[0039] The preset yield strain threshold is set to the yield strain limit value of the building material.

[0040] Preferably, when a normal data signal is generated, the current data transmission and reconstruction logic is maintained;

[0041] When a fake fault signal is generated, it is determined that the data link is faked or the sensor is faulty, and an alarm command is generated.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. This system achieves accurate quantification of transmission environment pressure by constructing a supply and demand coupling model; the system calculates the transmission pressure index by combining the frequency domain transformation characteristics of the original video stream with the channel status of the physical link; this mechanism can perceive the changes in communication load caused by dense fog, complex textures or network fluctuations in high-altitude operation scenarios in real time, providing a reliable physical basis for subsequent risk assessment, and effectively solving the problem that existing technologies lack quantitative assessment of transmission pressure in harsh communication environments;

[0044] 2. This system introduces a generative reconstruction risk assessment and feedback mechanism, which effectively suppresses the risk of visual hallucinations. The system comprehensively considers the natural decay of model confidence over time and the semantic mutation of feature vectors in space to calculate the hallucination index. This index can quantify in real time the risk of the generative artificial intelligence model producing images that do not conform to physical facts during the data missing repair process, preventing the monitoring screen from becoming disconnected from the actual construction status due to model inference bias, and ensuring the credibility of visual data.

[0045] 3. This system establishes an intelligent dual-modal switching strategy based on game theory. By performing quantitative matching analysis of the illusion index and transmission delay data, the system can make intelligent decisions between video stream transmission and topological feature stream transmission modes according to the current environment. This not only ensures rich details of the picture when the network is good, but also ensures the continuity of monitoring services and low-latency delivery of core structural data by forcibly switching to geometric feature transmission under extreme communication conditions.

[0046] 4. This system proposes a closed-loop verification method for physical consistency based on rigid body geometric invariants. Using Building Information Modeling (BIM) as a theoretical benchmark, the system calculates the structural topological consistency residual between the real-time measured length and the theoretical static length of the rigid connection edge, and compares it with the yield strain limit of the building material. This method can effectively eliminate environmental interference and accurately identify data forgery or sensor failure from a physical and mechanical perspective, solving the technical problem that traditional solutions cannot physically verify the authenticity of monitoring data. Attached Figure Description

[0047] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0048] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0050] Example 1:

[0051] Please see Figure 1 A real-time construction quality inspection system based on computer vision includes:

[0052] The data acquisition and management center is used to retrieve the original video stream data, channel status data, and transmission delay data from the high-altitude work surface, and send the original video stream data, channel status data, and transmission delay data to the transmission stress assessment unit.

[0053] The transmission stress assessment unit is used to perform transmission environment stress assessment and analysis on the raw video stream data and channel status data to obtain the transmission stress index.

[0054] The hallucination risk analysis unit is used to perform generative reconstruction risk assessment and feedback analysis based on the transmission pressure index to obtain the hallucination index;

[0055] The mode decision unit is used to perform game-theoretic transmission decision quantitative matching analysis on illusion index and transmission delay data to obtain mode selection discrimination value, and compare and analyze the mode selection discrimination value with the preset discrimination value threshold.

[0056] If the mode selection discrimination value is greater than or equal to the preset discrimination value threshold, a video stream transmission signal is generated;

[0057] If the mode selection discrimination value is less than the preset discrimination value threshold, a topology feature flow transmission signal is generated;

[0058] The consistency verification unit is used to perform physical consistency closed-loop verification analysis on the acquired heterogeneous data in response to the topological feature flow transmission signal, obtain the structural topological consistency residual, and compare and analyze the structural topological consistency residual with the preset yield strain threshold.

[0059] If the structural topology consistency residual is less than or equal to the preset yield strain threshold, a normal data signal is generated.

[0060] If the structural topology consistency residual is greater than the preset yield strain threshold, a fake fault signal is generated.

[0061] This embodiment provides a computer vision-based real-time construction quality inspection system, which is configured to solve the problem of video transmission and reconstruction quality verification in high-altitude operation environments;

[0062] The system mainly includes a data acquisition and management center, a transmission stress assessment unit, an illusion risk analysis unit, a pattern decision unit, and a consistency verification unit. The data acquisition and management center, as the system's sensing front end, is configured to retrieve raw video stream data, channel status data, and transmission delay data from the high-altitude work surface in real time. Raw video stream data is acquired using a high-resolution binocular stereo camera or an RGB-D depth camera; position and attitude data are acquired using an inertial measurement unit or an RTK positioning module; channel status data is directly read from the register status values ​​of the underlying network driver; and transmission delay data is obtained through timestamp comparison. The acquisition center synchronously sends these three types of data to the transmission stress assessment unit.

[0063] The transmission stress assessment unit is used to conduct transmission environment stress assessment and analysis on the original video stream data and channel status data, and calculates the transmission stress index by constructing a supply and demand coupling model. The hallucination risk analysis unit performs generative reconstruction risk assessment feedback analysis based on the transmission stress index, and calculates the hallucination index. This index is used to quantify the risk level of the generative AI model generating false images when data is missing. The mode decision unit performs game-theoretic transmission decision quantification and matching analysis. It receives the hallucination index and transmission delay data, calculates the mode selection discrimination value, and compares and analyzes the mode selection discrimination value with the preset discrimination value threshold.

[0064] If the mode selection discrimination value is greater than or equal to the preset discrimination value threshold, the system determines that the current environment is suitable for video transmission and generates a video stream transmission signal; if the mode selection discrimination value is less than the preset discrimination value threshold, the system determines that the video transmission risk is too high, generates a topology feature stream transmission signal, and forcibly switches to the geometric feature transmission mode; the consistency verification unit responds to the topology feature stream transmission signal, performs physical consistency closed-loop verification analysis on the collected heterogeneous data, calculates the structural topology consistency residual, and compares the residual with the preset yield strain threshold; if the residual is less than or equal to the threshold, a data normal signal is generated; if the residual is greater than the threshold, a fake fault signal is generated.

[0065] Through the above system architecture, this application realizes intelligent switching from simple video transmission to dual-modal transmission of visual and physical features, ensuring the physical authenticity of construction quality monitoring data in harsh communication environments.

[0066] Example 2:

[0067] The process of transmission environment stress assessment and analysis is as follows:

[0068] Frequency domain transformation analysis is performed on the raw video stream data to extract normalized visual entropy, and dynamic gain coefficients are constructed based on the normalized visual entropy.

[0069] Obtain the baseline bit rate required for the minimum recognition accuracy of machine vision, and multiply the baseline bit rate with the dynamic gain coefficient to obtain the data requirement.

[0070] The channel state data is parsed to extract the instantaneous detection bandwidth of the physical link and the channel packet loss rate;

[0071] Obtain the preset protocol overhead efficiency factor, and multiply the instantaneous probe bandwidth, channel packet loss rate, and protocol overhead efficiency factor to obtain the effective carrying capacity;

[0072] Dividing the data demand by the effective carrying capacity yields the transmission pressure index.

[0073] This embodiment elaborates on the specific logic of transmission environment stress assessment and analysis, aiming to quantify the coupling relationship between visual stream high-entropy noise and channel capacity;

[0074] The transmission stress assessment unit performs frequency domain transform analysis on the input raw video stream data; specifically, it performs block discrete cosine transform on each frame of the raw video stream and calculates all... The sum of the energy of the AC coefficients in an image patch, and the ratio of that energy to the total energy of the image; the formula is: ,in, High-frequency component energy, The total energy of the image; The range of values ​​is It is used to characterize the complexity of image texture and serves as the basic variable for calculating dynamic gain coefficient, reflecting the degree of reduction in coding efficiency caused by dense fog, rain, snow, or complex textures.

[0075] To determine the demand-side parameters for data transmission, the system obtains a preset baseline bit rate. The baseline bit rate This refers to the bit rate, measured in Mbps, required to achieve the minimum recognition accuracy of machine vision algorithms under standard low-entropy conditions; the system will use the baseline bit rate. With based The data demand at the current moment is calculated by multiplying the gain factors. ;

[0076] The transmission stress assessment unit analyzes the channel state data and extracts the instantaneous probe bandwidth of the physical link. and channel packet loss rate Simultaneously, obtain the preset protocol overhead efficiency factor. This factor is a constant used to correct the deviation between the theoretical bandwidth and the actual payload bandwidth; the system measures the instantaneous detection bandwidth. Channel packet loss rate and protocol overhead efficiency factor Perform multiplication to calculate the effective capacity of the current link. ;

[0077] Data demand Divide by effective carrying capacity The dimensionless transmission pressure index was calculated. The calculation formula is as follows:

[0078]

[0079] It should be noted that, in order to prevent division by zero errors when the channel is completely interrupted, the channel packet loss rate is included in the calculation process. Set a numerical saturation threshold, i.e., set a value. Or introduce a minimum value into the denominator. Make corrections;

[0080] in, These are the pre-calibrated weighting coefficients; regarding The calibration process in this embodiment uses an offline fitting method: collecting historical foggy video sample datasets and statistically analyzing the entropy values ​​of different samples. The incremental compression bitrate required to maintain recognition accuracy Calculate the bit rate increment Compared to the baseline bit rate The ratio of the sample entropy value to the sample entropy value is established. The linear relationship was determined, and dimensionless weighting coefficients were fitted using linear regression analysis. The value at this time; This formula represents the relative bitrate demand growth rate caused by the increase in unit visual entropy. It intuitively reflects the pressure of the current network environment on carrying high-complexity video streams through the supply-demand ratio, providing a physical basis for subsequent risk assessment.

[0081] Example 3:

[0082] The generative reconstruction risk assessment feedback analysis process is as follows:

[0083] Obtain the preset model confidence decay constant, and obtain the duration between the arrival time of the most recent fully validated keyframe and the current time, and set the duration as the keyframe interval duration;

[0084] The model confidence decay constant, transmission pressure exponent, and key frame interval duration are multiplied together, and the exponential decay value is calculated based on the product result. The natural decay component is obtained by subtracting the exponential decay value from the value 1.

[0085] Get the currently received sparse feature vector, retrieve the historical feature vector library, calculate the cosine distance between the sparse feature vector and the historical feature vector library, and set the cosine distance as the feature vector cosine distance.

[0086] Obtain the preset feature drift weight coefficients, and multiply the feature drift weight coefficients with the cosine distance of the feature vector to obtain the semantic mutation component;

[0087] The illusion index is obtained by summing the natural decay component and the semantic mutation component.

[0088] This embodiment describes in detail the process of generative reconstruction risk assessment and feedback analysis, which is used to quantify the risk of hallucinations when the receiving end uses a generative model to repair lost data;

[0089] The hallucination risk analysis unit obtains the preset model confidence decay constant. The constant The calibration method is as follows: adversarial testing is conducted on the generative model in a laboratory environment, with pure noise data as input, and the average time for the model to generate realistic but fake images is recorded. Then take Simultaneously, the system acquires the keyframe interval duration. That is, the time elapsed between the arrival time of the most recent fully validated keyframe and the current time.

[0090] The system's model confidence decay constant The aforementioned calculated transmission pressure index and keyframe interval duration Perform multiplication and calculate using the natural constant. The negative exponential function value with base is used to obtain the exponentially decaying value. Subtracting the exponential decay value from the numerical value 1 yields the natural decay component; this component characterizes the natural decline in the reliability of the model output as time progresses and transmission pressure increases.

[0091] Furthermore, to introduce content-based real-time verification, the system inputs the current frame into a pre-built lightweight convolutional neural network. This network is pre-trained in a self-supervised manner on a construction scene dataset using contrastive loss, exhibiting robustness to lighting changes while remaining sensitive to structural semantics. The extracted dimensions are... The semantic embedding vector is obtained and set as the currently received semantic feature vector; the local storage historical feature vector library is retrieved; the cosine distance between the current semantic feature vector and the corresponding feature in the historical feature vector library is calculated and set as the feature vector cosine distance. ; Obtain the preset feature drift weight coefficients ,Will and Perform product processing to obtain semantic mutation components;

[0092] The illusion index is obtained by summing the natural decay component and the semantic mutation component. Its mathematical expression is:

[0093]

[0094] This formula combines the credibility decay in the time dimension and the semantic mutation in the spatial dimension. The larger the value, the higher the likelihood that the reconstructed image deviates from physical reality, and the lower the reliability.

[0095] Example 4:

[0096] The quantitative matching analysis process of game transmission decision is as follows:

[0097] Obtain a preset hallucination index threshold, and subtract the hallucination index from the preset hallucination index threshold to obtain the reality difference value;

[0098] Obtain the preset authenticity utility weight, multiply the authenticity utility weight by the authenticity difference value, and obtain the authenticity utility value.

[0099] Obtain the maximum allowed latency limit for the task, divide the transmission latency data by the maximum latency limit, and obtain the normalized latency ratio.

[0100] Obtain the preset real-time penalty weight, and multiply the real-time penalty weight with the normalized latency ratio to obtain the real-time penalty value;

[0101] Subtracting the real-time penalty value from the real utility value yields the pattern selection discriminant value.

[0102] This embodiment details the process of quantitative matching analysis of game transmission decisions, which aims to balance data authenticity and transmission real-time performance.

[0103] The pattern decision unit obtains the preset hallucination index threshold. The threshold This is determined based on numerous subjective visual quality assessment tests, representing the maximum level of falsehood the system can tolerate; a preset hallucination index threshold is set. Subtract the currently calculated hallucination index Obtain the authenticity difference value; obtain the preset authenticity utility weight. ,Will The product of the actual utility value and the actual utility value is obtained.

[0104] At the same time, the system obtains the maximum allowable latency limit for the task. This value is set according to the security requirements of remote control; it includes the current transmission delay data. Divide by the maximum latency limit This yields the dimensionless normalized delay ratio; and the preset real-time penalty weight is obtained. ,Will The real-time penalty value is obtained by multiplying the normalized delay ratio;

[0105] The mode selection discriminant value is obtained by subtracting the real-time penalty value from the actual utility value, and then further subtracting the preset system switching bias constant from the result. The calculation formula is as follows:

[0106]

[0107] in, The default value is 1, used to adjust the conservatism of the system under critical conditions; here, the preset discriminant threshold is normalized to 0. This indicates that the current hallucination risks and latency are within a controllable range, and the benefits of transmitting the video stream outweigh the risks; when When this occurs, it indicates that the system can no longer maintain reliable video transmission and must switch modes.

[0108] Example 5:

[0109] The physical consistency closed-loop verification and analysis process is as follows:

[0110] Collect observable rigid connection edges within the current field of view, and obtain the real-time three-dimensional coordinate vectors of the first and second endpoints of the rigid connection edges.

[0111] Using the Euclidean distance operator, the distance between the real-time three-dimensional coordinate vectors of the first and second endpoints is calculated to obtain the real-time side length.

[0112] Retrieve the theoretical static length of the rigid connection edge in the preset building information model;

[0113] Subtract the theoretical static length from the real-time side length to obtain the length difference. Divide the length difference by the theoretical static length to obtain the unilateral strain deviation.

[0114] The unilateral strain deviation of all rigid connection edges within the field of view is averaged to obtain the structural topology consistency residual.

[0115] This embodiment illustrates the process of physical consistency closed-loop verification analysis, which is the core verification logic after the system switches to topology feature flow mode;

[0116] The consistency verification unit acquires observable rigid connection edges within the current field of view; using the binocular parallax principle or depth channel data, it back-projects the two-dimensional pixel coordinates to the three-dimensional camera coordinate system to obtain the real-time three-dimensional coordinate vector of the first endpoint of the rigid connection edge with physical scale. Real-time three-dimensional coordinate vector of the second endpoint Using the Euclidean distance operator, the distance between the two endpoint vectors is calculated to obtain the real-time side length. ;

[0117] Retrieve real-time six-DOF pose data of the camera recorded by the inertial measurement unit or RTK positioning module, and project all rigid connection edges in the building information model onto the two-dimensional imaging plane of the current camera using a perspective projection matrix; calculate the Euclidean distance between the rigid connection edges acquired in the current field of view and the projected BIM edges, select the projection edge with the smallest distance and direction angle less than the preset tolerance as the matching object, thereby locking the unique index ID of the edge in the BIM model; based on the index ID, retrieve the theoretical static length of the rigid connection edge in the preset building information model.

[0118] The system retrieves the theoretical static length of the rigid connection edge in the preset building information model. Real-time side length Subtract the theoretical static length Take the absolute value to obtain the length difference; divide this length difference by the theoretical static length. This yields the unilateral strain deviation of that side;

[0119] To eliminate single-point measurement errors, the system performs measurements on all points within the field of view. The unilateral strain deviation of the rigid connection edge is averaged to calculate the structural topology consistency residual. Its mathematical expression is:

[0120]

[0121] This method utilizes the principle of rigid body geometric invariants, which states that regardless of the overall displacement or rotation of a building due to strong winds, the length of rigid components should remain constant; therefore... It can effectively eliminate environmental interference and accurately reflect the physical state of the structure.

[0122] Example 6:

[0123] Structural topology consistency residuals are used to characterize the rate of change of internal geometric invariants of structural components.

[0124] The preset yield strain threshold is set to the yield strain limit value of the building material.

[0125] This embodiment clearly defines the physical meaning of the verification parameters; structural topology consistency residuals. Used to characterize the rate of change of internal geometric invariants of a structural component, rather than changes in absolute position; preset yield strain threshold. This is set as the yield strain limit value for building materials; for example, for Q345 steel, its yield strain limit value is typically approximately [value missing]. Using this physical limit as the discrimination threshold ensures that the alarm signal has a clear material mechanics basis.

[0126] Example 7:

[0127] When a normal data signal is generated, the current data transmission and reconstruction logic is maintained;

[0128] When a fake fault signal is generated, it is determined that the data link is faked or the sensor is faulty, and an alarm command is generated.

[0129] At the same time, the sparse feature vector extracted at the current moment is updated to the historical feature vector library, and the historical feature vector with the earliest timestamp is removed, so as to adapt to the dynamic changes in the building form during construction through the sliding window mechanism and prevent normal semantic drift caused by construction progress from being misjudged as hallucination risk.

[0130] This embodiment describes the system response logic based on the verification result; when a normal data signal is generated, it indicates that... At this point, the monitored object is determined to be in a physically safe state, and the system maintains the current data transmission and reconstruction logic; when a fake fault signal is generated, it indicates that... If the system determines that the data link is forged or the sensor has malfunctioned, it will immediately generate an alarm command to prompt the operator to intervene manually; this mechanism constitutes the last line of defense for the system.

[0131] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A real-time construction quality inspection system based on computer vision, characterized in that, include: The data acquisition and management center is used to retrieve the original video stream data, channel status data, and transmission delay data from the high-altitude work surface, and send the original video stream data, channel status data, and transmission delay data to the transmission stress assessment unit. The transmission stress assessment unit is used to perform transmission environment stress assessment and analysis on the raw video stream data and channel status data to obtain the transmission stress index. The hallucination risk analysis unit is used to perform generative reconstruction risk assessment and feedback analysis based on the transmission pressure index to obtain the hallucination index; The mode decision unit is used to perform game-theoretic transmission decision quantitative matching analysis on illusion index and transmission delay data to obtain mode selection discrimination value, and compare and analyze the mode selection discrimination value with the preset discrimination value threshold. If the mode selection discrimination value is greater than or equal to the preset discrimination value threshold, a video stream transmission signal is generated; If the mode selection discrimination value is less than the preset discrimination value threshold, a topology feature flow transmission signal is generated; The consistency verification unit is used to perform physical consistency closed-loop verification analysis on the acquired heterogeneous data in response to the topological feature flow transmission signal, obtain the structural topological consistency residual, and compare and analyze the structural topological consistency residual with the preset yield strain threshold. If the structural topology consistency residual is less than or equal to the preset yield strain threshold, a normal data signal is generated. If the structural topology consistency residual is greater than the preset yield strain threshold, a fake fault signal is generated.

2. The real-time construction quality detection system based on computer vision according to claim 1, characterized in that, The transmission environment stress assessment and analysis process is as follows: Frequency domain transformation analysis is performed on the raw video stream data to extract normalized visual entropy, and dynamic gain coefficients are constructed based on the normalized visual entropy. Obtain the baseline bit rate required for the minimum recognition accuracy of machine vision, and multiply the baseline bit rate with the dynamic gain coefficient to obtain the data requirement. The channel state data is parsed to extract the instantaneous detection bandwidth of the physical link and the channel packet loss rate; Obtain the preset protocol overhead efficiency factor, and multiply the instantaneous probe bandwidth, channel packet loss rate, and protocol overhead efficiency factor to obtain the effective carrying capacity; Dividing the data demand by the effective carrying capacity yields the transmission pressure index.

3. The real-time construction quality detection system based on computer vision according to claim 1, characterized in that, The generative reconstruction risk assessment feedback analysis process is as follows: Obtain the preset model confidence decay constant, and obtain the duration between the arrival time of the most recent fully validated keyframe and the current time, and set the duration as the keyframe interval duration; The model confidence decay constant, transmission pressure exponent, and key frame interval duration are multiplied together, and the exponential decay value is calculated based on the product result. The natural decay component is obtained by subtracting the exponential decay value from the value 1. Get the currently received sparse feature vector, retrieve the historical feature vector library, calculate the cosine distance between the sparse feature vector and the historical feature vector library, and set the cosine distance as the feature vector cosine distance. Obtain the preset feature drift weight coefficients, and multiply the feature drift weight coefficients with the cosine distance of the feature vector to obtain the semantic mutation component; The illusion index is obtained by summing the natural decay component and the semantic mutation component.

4. The real-time construction quality detection system based on computer vision according to claim 1, characterized in that, The quantitative matching analysis process for game transmission decisions is as follows: Obtain a preset hallucination index threshold, and subtract the hallucination index from the preset hallucination index threshold to obtain the reality difference value; Obtain the preset authenticity utility weight, multiply the authenticity utility weight by the authenticity difference value, and obtain the authenticity utility value. Obtain the maximum allowed latency limit for the task, divide the transmission latency data by the maximum latency limit, and obtain the normalized latency ratio. Obtain the preset real-time penalty weight, and multiply the real-time penalty weight with the normalized latency ratio to obtain the real-time penalty value; Subtracting the real-time penalty value from the real utility value yields the pattern selection discriminant value.

5. A real-time construction quality detection system based on computer vision according to claim 1, characterized in that, The physical consistency closed-loop verification and analysis process is as follows: Collect observable rigid connection edges within the current field of view, and obtain the real-time three-dimensional coordinate vectors of the first and second endpoints of the rigid connection edges. Using the Euclidean distance operator, the distance between the real-time three-dimensional coordinate vectors of the first and second endpoints is calculated to obtain the real-time side length. Retrieve the theoretical static length of the rigid connection edge in the preset building information model; Subtract the theoretical static length from the real-time side length to obtain the length difference. Divide the length difference by the theoretical static length to obtain the unilateral strain deviation. The unilateral strain deviation of all rigid connection edges within the field of view is averaged to obtain the structural topology consistency residual.

6. The real-time construction quality detection system based on computer vision according to claim 1, characterized in that, The structural topology consistency residual is used to characterize the rate of change of the internal geometric invariants of the structural component. The preset yield strain threshold is set to the yield strain limit value of the building material.

7. A real-time construction quality detection system based on computer vision according to claim 1, characterized in that, When a normal data signal is generated, the current data transmission and reconstruction logic is maintained; When a fake fault signal is generated, it is determined that the data link is faked or the sensor is faulty, and an alarm command is generated.