A safe production video monitoring system

Abstract

The application discloses a kind of safety production video monitoring systems, specifically related to video monitoring system field, comprising: video data acquisition module, image analysis module, correction module, comprehensive evaluation module and dynamic adjustment module;Video data acquisition module real-time acquisition video data in production operation process, and upload to image analysis module;Image analysis module extracts the color feature of target, contour feature, motion feature, constructs risk feature matrix, and generates image error marker coefficient;Correction module is used to correct the image error marker coefficient and handle;Comprehensive evaluation module is used to carry out comprehensive evaluation to the image error marker coefficient after correction, generates evaluation adjustment coefficient;Dynamic adjustment module is used to carry out secondary correction to the image error marker coefficient according to the evaluation adjustment coefficient generated, the risk identification precision of the present application is significantly improved, reduces false alarm and miss rate, while enhancing the adaptability and reliability to different production scenes.

Description

Technical Field

[0001] This invention relates to the field of video surveillance systems, and more specifically, to a video surveillance system for safe production. Background Technology

[0002] Against the backdrop of industrial production automation and intelligent upgrading, video surveillance systems for safety production have become a core technological means for risk prevention and control in high-risk industries such as chemical, mining, and construction. However, current mainstream monitoring systems still suffer from multi-dimensional technical bottlenecks, making it difficult to meet the precise safety management needs of complex production scenarios.

[0003] Traditional video surveillance relies too heavily on manual monitoring. Faced with massive real-time video streams, monitoring personnel are prone to missing risk assessments due to fatigue. Furthermore, manual analysis is inefficient, often relying solely on motion detection or simple contour recognition, which fails to capture key risk features such as abnormal color changes in equipment (e.g., discoloration caused by high temperatures) or personnel wearing inappropriate clothing, resulting in a one-sided approach to risk identification.

[0004] The interference from complex industrial environments further exacerbates the problem of insufficient recognition accuracy. Factors such as light fluctuations, mechanical vibrations, and dust obstruction in production sites can significantly increase video feature noise. Traditional systems lack effective noise suppression mechanisms, resulting in a high feature extraction error rate. Furthermore, existing systems often use fixed thresholds for risk assessment, failing to consider the inherent risk differences in different scenarios such as chemical plants and mines. In high-risk scenarios, overly lenient threshold settings can lead to missed detections, while in conventional scenarios, overly strict thresholds can generate numerous false positives.

[0005] Traditional monitoring systems operate independently, with feature analysis, error correction, and risk assessment processes running in parallel. The correction process fails to differentiate between characteristic deviation patterns, employing a uniform approach for different types of deviations, such as left-hand or right-hand tilt, resulting in insufficient accuracy in error quantification. Furthermore, the assessment phase lacks spatiotemporal correlation analysis capabilities, making it difficult to identify rapid spread of equipment failures or localized risk clustering. These shortcomings significantly hinder the adaptability and reliability of existing systems when dealing with complex and ever-changing production risks.

[0006] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0007] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a video surveillance system for safe production to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a video surveillance system for safe production, comprising: a video data acquisition module, an image analysis module, a correction module, a comprehensive evaluation module, and a dynamic adjustment module;

[0009] The video data acquisition module collects video data in real time during the production process and uploads it to the image analysis module;

[0010] The image analysis module extracts the target's color features, contour features, and motion features, constructs a risk feature matrix, and generates image error labeling coefficients;

[0011] The correction module is used to correct the image error marker coefficients;

[0012] The comprehensive evaluation module is used to comprehensively evaluate the corrected image error labeling coefficients and generate evaluation adjustment coefficients;

[0013] The dynamic adjustment module is used to perform secondary correction on the image error labeling coefficients based on the generated evaluation adjustment coefficients.

[0014] In a preferred embodiment, the operation of the image analysis module includes the following:

[0015] The extracted color features, contour features, and motion features are vectorized and encoded to form an initial feature vector set;

[0016] The initial feature vector set is reorganized into a matrix according to the time series and spatial distribution dimensions to construct the risk feature matrix F, where the row dimension of the risk feature matrix corresponds to the time series of video frames, and the column dimension corresponds to the vector dimension of multi-dimensional features.

[0017] The initial risk feature distribution is defined by parameter ω0. In the i-th iteration, the rendering function g(ω) is used. i c) Reconstruct and render the risk feature matrix F to obtain the reconstructed feature matrix. Where c represents the camera parameters corresponding to the video capture viewpoint;

[0018] Using a pre-trained noise prediction network, the reconstructed feature matrix is ​​obtained from the input. Predicted time-time noise λ;

[0019] The noise distribution s obtained by minimizing the noise prediction network sampling λ (λ|c A ,c B ) and the true noise distribution g λ (λ|c A ,c B The risk feature matrix F is optimized using the KL divergence between the two features, as shown in the following formula:

[0020]

[0021] And based on the optimization process, image error labeling coefficients λ are generated. i It is used to quantify the degree of deviation between the current video frame features and the standard compliance features.

[0022] In a preferred embodiment, the operation of the correction module includes the following:

[0023] In the statistical risk feature matrix, the number of left-hand sampling points M corresponding to the current video frame's features within the standard compliance feature space. l Number of sampling points in the middle section M d Number of sampling points on the right side M r The statistical formula is:

[0024]

[0025] Where M is the number of sampling points, λ t , is the fitting center parameter of the standard compliance feature, corresponding to the baseline center value of the color feature, contour feature, and motion feature. K1 is the fitting sampling interval. X is the left boundary parameter value of the current feature in the risk feature matrix when calculating the number of sampling points on the left, the right boundary parameter value when calculating the number of sampling points on the right, and the middle boundary parameter value when calculating the number of sampling points in the middle.

[0026] Determine the characteristic deviation pattern based on the percentage data:

[0027] If M d ≥0.75, is determined to be an unbiased equilibrium mode;

[0028] If M l >M r And M l ≥0.4 indicates a left-leaning bias pattern;

[0029] If M r >M l And M r ≥0.4 indicates a right-leaning deviation pattern;

[0030] Combining deviation patterns and scene risk weights, the image error labeling coefficients λ are... i Perform calibration operation:

[0031] Retrieve the risk weighting factor W for the production scenario r ;

[0032] In the balanced unbiased mode, the error labeling coefficient after calibration is:

[0033]

[0034] In left-tilt bias mode, the calibration error marking coefficient is:

[0035]

[0036] In right-tilt bias mode, the calibration error marking coefficient is:

[0037]

[0038] In a preferred embodiment, the operation of the comprehensive evaluation module includes the following:

[0039] For color features, contour features, and motion features, based on the real-time risk situation in the production scenario, feature weights are dynamically assigned. The weights for color features, contour features, and motion features are labeled w1, w2, and w3, respectively, and the weight assignment rules are as follows:

[0040] w1 = 0.2 + 0.1 × L;

[0041] w2 = 0.5 - 0.1 × L;

[0042] w3 = 0.3;

[0043] Where L is the scenario risk level, with a value range of [1, 5], and w1+w2+w3=1.

[0044] In a preferred embodiment, the corrected image error labeling coefficients are... Perform spatiotemporal correlation analysis and calculate the rate of change of the error coefficients within the window.

[0045] If Δλ ≥ 0.02 / s, it is judged as a rapidly changing risk;

[0046] Calculate the mean error coefficient of each sub-region in a 3×3 grid. If the average value of a certain sub-region exceeds 1.5 times the overall average value, it is judged as a local clustering risk.

[0047] In a preferred embodiment, an evaluation adjustment coefficient β is generated based on the spatiotemporal correlation results and dynamic feature weights. If only the risk of rapid change exists, then:

[0048]

[0049] If only local clustering risk exists, then:

[0050]

[0051] If both risks exist simultaneously, then:

[0052]

[0053] Finally, β = max(β1, β2, β3) was chosen as the evaluation adjustment coefficient for the output.

[0054] In a preferred embodiment, the operation of the dynamic adjustment module includes the following:

[0055] Based on the output evaluation adjustment coefficient β and the risk level threshold range of the production scenario, a correction threshold E is dynamically generated;

[0056] E = 0.8β (Scenario risk level L∈[1,2);

[0057] E = β (Scenario risk level L∈[2,4);

[0058] E = 1.2β (Scenario risk level L∈[4,5);

[0059] Statistically corrected error coefficients Calculate the feedback adjustment factor δ using historical data:

[0060]

[0061] In the formula, n represents the amount of data.

[0062] In a preferred embodiment, the error labeling coefficients of the corrected image are adjusted by combining the correction threshold E and the feedback adjustment factor δ. Perform a second correction, and finally output the second correction coefficient.

[0063]

[0064] Where w represents the weights of color features, contour features, and motion features, used to differentiate and amplify the error impact of different features.

[0065] The technical effects and advantages of the video surveillance system for safe production of this invention are as follows:

[0066] 1. This invention constructs a risk identification system that integrates color, contour, and motion features through an image analysis module. Combined with a pre-trained noise prediction network and KL divergence optimization algorithm, it can effectively filter on-site interference such as light fluctuations and dust obstruction. It quantifies the deviation between risk features and standard compliance features into precise image error labeling coefficients, thereby improving the accuracy of automated identification and avoiding risk omissions caused by human fatigue.

[0067] 2. This invention distinguishes between three types of deviation modes—unbiased equilibrium, left-leaning, and right-leaning—through a correction module, and achieves differentiated calibration by combining scenario risk weights, thus avoiding error amplification caused by correction. Simultaneously, the comprehensive evaluation module dynamically allocates feature weights based on real-time risk levels and captures rapidly changing and locally clustered risks through spatiotemporal correlation analysis. The dynamic adjustment module calculates feedback adjustment factors through a risk level and correction threshold linkage mechanism, combined with historical error data, to compensate for the impact of short-term fluctuations such as instantaneous light interference on the correction results. Furthermore, the introduction of feature-differentiated weights ensures that the secondary correction coefficient aligns with long-term risk trends. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the structure of a video surveillance system for safe production according to the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0070] Example 1

[0071] Figure 1 The present invention provides a video surveillance system for safe production, comprising: a video data acquisition module, an image analysis module, a correction module, a comprehensive evaluation module, and a dynamic adjustment module;

[0072] The video data acquisition module collects video data in real time during the production process and uploads it to the image analysis module;

[0073] The image analysis module extracts the target's color features, contour features, and motion features, constructs a risk feature matrix, and generates image error labeling coefficients;

[0074] The correction module is used to correct the image error marker coefficients;

[0075] The comprehensive evaluation module is used to comprehensively evaluate the corrected image error labeling coefficients and generate evaluation adjustment coefficients;

[0076] The dynamic adjustment module is used to perform secondary correction on the image error labeling coefficients based on the generated evaluation adjustment coefficients.

[0077] The operation of the image analysis module includes the following:

[0078] The extracted color features, contour features, and motion features are vectorized and encoded to form an initial feature vector set;

[0079] The initial feature vector set is reorganized into a matrix according to the time series and spatial distribution dimensions to construct the risk feature matrix F, where the row dimension of the risk feature matrix corresponds to the time series of video frames, and the column dimension corresponds to the vector dimension of multi-dimensional features.

[0080] The initial risk feature distribution is defined by parameter ω0. In the i-th iteration, the rendering function g(ω) is used. i c) Reconstruct and render the risk feature matrix F to obtain the reconstructed feature matrix. Where c represents the camera parameters corresponding to the video capture viewpoint;

[0081] Using a pre-trained noise prediction network, the reconstructed feature matrix is ​​obtained from the input. Predicted time-time noise λ;

[0082] The noise distribution s obtained by minimizing the noise prediction network sampling λ (λ|c A ,c B ) and the true noise distribution g λ (λ|c A ,c B The risk feature matrix F is optimized using the KL divergence between the two features, as shown in the following formula:

[0083]

[0084] And based on the optimization process, image error labeling coefficients λ are generated. i It is used to quantify the degree of deviation between the current video frame features and the standard compliance features.

[0085] By extracting and quantifying three core features—color, contour, and motion—and combining time series and spatial distribution to construct a risk feature matrix, we can comprehensively capture the risk correlation features of personnel, equipment, and materials in production scenarios, avoiding the one-sidedness of single feature analysis. At the same time, by optimizing the risk feature matrix through KL divergence minimization, we can transform the deviation between the current video frame features and standard compliance features into quantifiable image error labeling coefficients, making the characterization of risk differences more objective.

[0086] The pre-trained noise prediction network can effectively filter feature noise caused by common interference factors in production sites, such as light fluctuations and equipment vibrations. The reconstruction rendering process combined with camera parameters can adapt to the differences in feature acquisition under different monitoring perspectives, reducing the impact of perspective deviation on feature analysis. Through iterative optimization, the gap between the predicted noise distribution and the real noise distribution is continuously narrowed, so that the generated image error labeling coefficients can dynamically fit scene changes, significantly reducing the misjudgment rate caused by environmental complexity and improving the system's stable operation capability in diverse production scenarios.

[0087] The operation of the correction module includes the following:

[0088] In the statistical risk feature matrix, the number of left-hand sampling points M corresponding to the current video frame's features within the standard compliance feature space. l Number of sampling points in the middle section M d Number of sampling points on the right side M r The statistical formula is:

[0089]

[0090] Where M is the number of sampling points, λ t , is the fitting center parameter of the standard compliance feature, corresponding to the baseline center value of the color feature, contour feature, and motion feature. K1 is the fitting sampling interval. X is the left boundary parameter value of the current feature in the risk feature matrix when calculating the number of sampling points on the left, the right boundary parameter value when calculating the number of sampling points on the right, and the middle boundary parameter value when calculating the number of sampling points in the middle.

[0091] Determine the characteristic deviation pattern based on the percentage data:

[0092] If M d ≥0.75, is determined to be an unbiased equilibrium mode;

[0093] If M l >M r And M l ≥0.4 indicates a left-leaning bias pattern;

[0094] If M r >M l And M r ≥0.4 indicates a right-leaning deviation pattern;

[0095] Combining deviation patterns and scene risk weights, the image error labeling coefficients λ are... i Perform calibration operation:

[0096] Retrieve the risk weighting factor W for the production scenario r ;

[0097] In the balanced unbiased mode, the error labeling coefficient after calibration is:

[0098]

[0099] In left-tilt bias mode, the calibration error marking coefficient is:

[0100]

[0101] In right-tilt bias mode, the calibration error marking coefficient is:

[0102]

[0103] By statistically analyzing the number and proportion of sampling points, we can accurately distinguish between three characteristic deviation modes: unbiased equilibrium, left-tilt deviation, and right-tilt deviation. This avoids the drawbacks of traditional uniform correction methods that treat different deviation types in the same way. Furthermore, by combining scene risk weights and customized calibration formulas, we can adaptively adjust the image error labeling coefficients for different deviation modes. For example, the right-tilt deviation scenario has a higher risk, so the coefficient amplification is more prominent after calibration, enabling the error coefficients to more accurately reflect the severity of the actual risk and reduce misjudgment or omission of risk.

[0104] The correction module introduces a risk weighting factor for production scenarios, flexibly adjusting the weighting ratio of calibration based on the inherent risk differences in different scenarios such as chemical plants and mines. At the same time, sampling point statistics and deviation judgment are based on real-time video frames, which can quickly capture dynamic changes in features and allow the error labeling coefficient to adapt to the instantaneous risk fluctuations in the production site in a timely manner.

[0105] The operation of the comprehensive evaluation module includes the following:

[0106] For color features, contour features, and motion features, based on the real-time risk situation in the production scenario, feature weights are dynamically assigned. The weights for color features, contour features, and motion features are labeled w1, w2, and w3, respectively, and the weight assignment rules are as follows:

[0107] w1 = 0.2 + 0.1 × L;

[0108] w2 = 0.5 - 0.1 × L;

[0109] w3 = 0.3;

[0110] Where L is the scenario risk level, with a value range of [1, 5], and w1+w2+w3=1.

[0111] Error labeling coefficients for the corrected image Perform spatiotemporal correlation analysis and calculate the rate of change of the error coefficients within the window.

[0112] If Δλ ≥ 0.02 / s, it is judged as a rapidly changing risk;

[0113] Calculate the mean error coefficient of each sub-region in a 3×3 grid. If the average value of a certain sub-region exceeds 1.5 times the overall average value, it is judged as a local clustering risk.

[0114] Based on the spatiotemporal correlation results and dynamic feature weights, an evaluation adjustment coefficient β is generated. If only the risk of rapid change exists, then:

[0115]

[0116] If only local clustering risk exists, then:

[0117]

[0118] If both risks exist simultaneously, then:

[0119]

[0120] Finally, β = max(β1, β2, β3) was chosen as the evaluation adjustment coefficient for the output.

[0121] By associating the scenario risk level L with the weights of color, outline, and motion features, the weighting of key features such as abnormal equipment outlines and material movement trajectories can be increased in high-risk scenarios, and the weighting of personnel movement features can be increased in densely populated scenarios. This makes the assessment adjustment coefficient more focused on the core risk points of the scenario, avoiding the key risk identification bias caused by general weights. Spatiotemporal correlation analysis enables accurate assessment of risk from both dynamic trends and spatial distribution. In the time dimension, rapidly changing risks are identified through the rate of change of error coefficients, timely capturing instantaneous risks such as sudden equipment failures and personnel violations. In the spatial dimension, local clustered risks are identified through statistical analysis of 3×3 grid sub-regions, accurately locating specific areas where risks are concentrated. This dual-dimensional analysis allows the assessment adjustment coefficient to reflect both the dynamic evolution of risks and the spatial clustering characteristics of risks, significantly reducing the probability of missed or misjudged assessments compared to single-dimensional assessments.

[0122] The dynamic adjustment module's operation process includes the following:

[0123] Based on the output evaluation adjustment coefficient β and the risk level threshold range of the production scenario, a correction threshold E is dynamically generated;

[0124] E = 0.8β (Scenario risk level L∈[1,2);

[0125] E = β (Scenario risk level L∈[2,4);

[0126] E = 1.2β (Scenario risk level L∈[4,5);

[0127] Statistically corrected error coefficients Calculate the feedback adjustment factor δ using historical data:

[0128]

[0129] In the formula, n represents the amount of data.

[0130] By combining the correction threshold E and the feedback adjustment factor δ, the error labeling coefficients of the corrected image are... Perform a second correction, and finally output the second correction coefficient.

[0131]

[0132] Where w represents the weights of color features, contour features, and motion features, used to differentiate and amplify the error impact of different features.

[0133] By linking the assessment adjustment coefficient with a preset risk level threshold range, a correction threshold of 1.2 times the assessment adjustment coefficient is used in high-risk scenarios (such as chemical leak early warnings) to amplify the warning effect of the error coefficient and ensure that serious risks are not overlooked. In low-risk scenarios (such as routine mechanical inspections), a threshold of 0.8 times is used to suppress false alarms caused by slight fluctuations and avoid invalid early warnings interfering with operation and maintenance. This secondary correction not only conforms to the risk characteristics of the scenario but also accurately balances risk capture and false alarm control.

[0134] The feedback adjustment factor is calculated based on historical data of the corrected error label coefficient. It can compensate for short-term data fluctuations (such as error jumps caused by instantaneous light interference), making the secondary correction coefficient more in line with long-term risk trends and reducing the impact of accidental factors. By differentiating the weights of color, outline, and motion features (such as increasing the weight of motion features in scenarios of personnel violations), the error impact of core risk features can be amplified in a targeted manner, avoiding non-critical features from interfering with risk judgment, so that the final output secondary correction coefficient is both stable and reliable.

[0135] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0136] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0137] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0138] In the several embodiments provided in this application, it should be understood that the disclosed systems and devices can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0139] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0140] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0141] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps described in the various embodiments of this application. The aforementioned storage medium includes 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.

[0142] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0143] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A video monitoring system for safe production, characterized in that, Comprise: Video data acquisition module, image analysis module, correction module, comprehensive evaluation module and dynamic adjustment module; The video data acquisition module collects video data in real time during the production operation process, and uploads it to the image analysis module; The image analysis module extracts the color features, contour features and motion features of the target, constructs a risk feature matrix, and generates an image error marking coefficient; The correction module is used for correcting the image error marking coefficient; The comprehensive evaluation module is used for comprehensive evaluation of the corrected image error marking coefficient, and generates an evaluation adjustment coefficient; The dynamic adjustment module is used for secondary correction of the image error marking coefficient according to the generated evaluation adjustment coefficient; The running process of the image analysis module includes the following contents: The extracted color features, contour features and motion features are vectorized and coded to form an initial feature vector set; The initial feature vector set is matrixed and reorganized according to time sequence and spatial distribution dimensions to construct a risk feature matrix wherein a row dimension of the risk feature matrix corresponds to a time sequence of video frames, and a column dimension corresponds to a vector dimension of multi-dimensional features The initial risk characteristic distribution is defined by parameters. It indicates that in the first In the next iteration, through the rendering function Risk feature matrix Perform reconstruction rendering to obtain the reconstructed feature matrix. ,in These are the camera parameters corresponding to the video capture angle; Using the pre-trained noise prediction network, a reconstructed feature matrix is input Predicted noise at the time ; By minimizing the KL divergence between the noise distribution predicted by the noise prediction network and the noise distribution sampled from the noise distribution The risk feature matrix is optimized by minimizing the KL divergence between the real noise distribution , as follows:​ ； and generating image error indicia coefficients based on the optimization process for quantifying the degree of deviation of the current video frame characteristics from the standard compliant characteristics; The running process of the correction module includes the following contents: In the statistical risk feature matrix, the number of left sampling points of the current video frame corresponding feature in the standard compliance feature space , the number of middle sampling points , the number of right sampling points , and the statistical formula is: ； wherein, is the number of sampling points, is the fitting center parameter of the standard compliance feature, corresponding to the reference center value of the color feature, the contour feature, and the motion feature, is the fitting sampling interval, is the left boundary parameter value of the current feature in the risk feature matrix when calculating the number of left sampling points, is the right boundary parameter value when calculating the number of right sampling points, and is the middle boundary parameter value when calculating the number of middle sampling points; Determine the feature deviation mode according to the proportion data: If , the equalization unbiased mode is determined; If > and , the left deviation mode is determined. If < / < / < / and the right skew deviation pattern is determined. Combining bias patterns with scene risk weights to image error labeling coefficients Performing a calibration operation: Retrieving risk weight factors for production scenarios ; In the balanced unbiased mode, the calibrated error marking coefficient is: ； In the left deviation mode, the calibrated error marking coefficient is: ； In the right deviation mode, the calibrated error marking coefficient is: 。 2. The safety production video monitoring system according to claim 1, characterized in that: The running process of the comprehensive evaluation module includes the following contents: For color features, contour features, motion features, based on the real-time risk situation of the production scene, the feature weights are dynamically allocated, and the weights of color features, contour features and motion features are marked as , and respectively, and the weight allocation rule is: ； ； ； wherein, is a scene risk level, and the value range is , and .

3. The safety production video monitoring system according to claim 2, characterized in that: error flag coefficient for the modified image spatiotemporal correlation analysis is performed to calculate the rate of change of error coefficients within the window ; If , the rapid change risk is determined; Statistical 3x3 grid of each sub-region error coefficient mean If the mean of a sub-region exceeds 1.5 times the overall mean, it is determined that there is a risk of local clustering.

4. The safety production video monitoring system according to claim 3, characterized in that: According to the spatiotemporal correlation result and the dynamic characteristic weight, an evaluation adjustment coefficient is generated If only the rapid change risk exists, then: ； If there is only local aggregation risk, then: ； If there are two risks at the same time, then: ； Final selection Evaluation adjustment factor as output.

5. The safety production video monitoring system according to claim 4, characterized in that: The running process of the dynamic adjustment module includes the following contents: Adjusting the coefficients according to the evaluation of the output The dynamic generation of the correction threshold is based on the risk level threshold interval of the production scene ; ； ； ； Statistical correction error flag coefficient of historical data, calculating a feedback adjustment factor : ； Where n is the data volume.

6. The safety production video monitoring system according to claim 5, characterized in that: Combining the modified threshold with the feedback adjustment factor , the modified image error flag coefficient performing secondary modification, finally outputting the secondary modified coefficient : ； wherein, are color feature, contour feature, motion feature weights, used to differentiate the error impact of different features.

Images