Water conservancy construction safety supervision system and method based on artificial intelligence

Abstract

The invention discloses a water conservancy construction safety supervision system and method based on artificial intelligence, particularly relates to the field of water conservancy construction monitoring, is used for solving the problems of light noise interference and moving target detection, and utilizes frequency domain sub-band energy analysis and multi-scale decomposition to respectively generate an illumination fluctuation index and a noise stability index. The quantification of environment illumination and noise interference is realized; illumination and noise events are fused through time delay kernel matching, a differential threshold value is dynamically adjusted, switching between a high light noise sensitive mode and a low light noise stable mode is achieved, and the accuracy of moving target detection is improved; in combination with a water conservancy tool morphological template and multi-frame track cross validation, noise and reflective false targets are effectively eliminated, and a motion area is extracted; and then a sensor driving factor is generated through multi-source matching, the background updating rate is regulated and controlled, pixel degradation caused by mechanical vibration is inhibited, and the early warning stability of a monitoring system is improved.

Description

Technical Field

[0001] The present invention relates to the field of water conservancy construction monitoring. More specifically, the present invention relates to a water conservancy construction safety supervision system and method based on artificial intelligence. Background Art

[0002] The operating environment of water conservancy project construction is complex and changeable. The video monitoring system in the open-air scene is vulnerable to multi-source noise interference caused by sudden changes in sunlight, uneven infrared fill light at night, and vibration of construction machinery, resulting in difficulty for traditional background modeling methods to stably distinguish real moving targets from light and shadow disturbances. Especially under extreme conditions such as heavy rain and dust storms, the coupled effect of high-frequency noise and dynamic light intensifies the pixel-level dynamic threshold drift, leading to frequent false detections and missed detections, directly affecting the reliability of construction safety monitoring and the accuracy of real-time early warning.

[0003] Currently, most mainstream background subtraction algorithms use fixed or linearly adjusted thresholds to cope with light changes, without quantifying the spatio-temporal correlation of light noise disturbance events, resulting in lagging or overly sensitive threshold responses in mutation scenarios. For example, although the frame difference method based on a statistical model can dynamically update the background, it lacks cross-scale analysis of the stability of multi-band noise and cannot eliminate edge artifacts caused by high-frequency vibration; while the mixture Gaussian model can handle gradual light changes, but due to ignoring the time-sequence drift of infrared fill light caused by the periodic operation of construction equipment, it causes event mis-matching. In addition, existing methods rely on the self-consistency of visual data and do not fuse the physical characteristics verification of vibration sensors. The model update mechanism is disconnected from the construction risk level and the operating conditions of machinery, and is prone to degradation due to cumulative errors during long-term monitoring, making it difficult to adapt to the dynamic supervision requirements of water conservancy scenarios.

[0004] To solve the above problems, a technical solution is provided now. Summary of the Invention

[0005] To overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a water conservancy construction safety supervision system and method based on artificial intelligence. By dynamically compensating for the periodic fluctuations of infrared fill light, accurately screening light mutation and noise sudden increase events, and using a time-delay kernel function to match and calculate the event correlation, the efficient fusion of light and noise information is achieved; further, by generating a sensor-driven weight correction factor using the vibration spectrum data of construction equipment and the material transportation path data, dynamically adjusting the background model update rate, and suppressing pixel degradation caused by mechanical vibration, thereby ensuring the stability and adaptability of the background model during long-term monitoring. to solve the problems raised in the above background art.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A water conservancy construction safety supervision method based on artificial intelligence, including the steps:

[0008] S1: Perform frequency-domain subband energy analysis and directionally controllable multi-scale decomposition on the real-time monitoring video frames, respectively extract the low-frequency energy change rate to generate a time-varying illumination fluctuation index, and generate a multi-scale noise stability index through the high-frequency subband stability coefficient to quantify the illumination dynamic characteristics and noise interference level;

[0009] S2: Calculate the event correlation density by performing time-delay kernel function matching on the illumination event sequence and noise event sequence after dynamic compensation, and use the multi-scale noise stability index to dynamically adjust the noise determination threshold and matching normalization, and finally generate a global fusion selection coefficient;

[0010] S3: By comparing the global fusion selection coefficient with the critical value, use an exponential response or illumination variance attenuation function to adjust the basic difference threshold respectively, and achieve dynamic switching between the high-light-noise sensitive mode and the low-light-noise stable mode to optimize the moving target detection;

[0011] S4: Perform a difference operation on the current frame and the background model using the adjusted difference threshold, screen the connected regions in combination with the preset morphological template of water conservancy tools, and eliminate instantaneous noise and reflection pseudo-targets through multi-frame trajectory cross-validation, and output the moving target area;

[0012] S5: Perform multi-source matching on the moving target trajectory, vibration spectrum and transportation path, generate a weight correction factor driven by the sensor according to the matching result, dynamically adjust the background model update rate and suppress pixel degradation caused by mechanical vibration.

[0013] In a preferred embodiment, step S1 includes the following contents:

[0014] The specific acquisition logic of the time-varying illumination fluctuation index is as follows:

[0015] The input video frame is denoted as F(t), and its size is M×N. Use the directionally controllable multi-scale decomposition operator to process the input video frame to obtain several low-frequency subbands, denoted as: L α (t,u,v)(α = 1,2,…,S L ); where L α (t,u,v) represents the pixel value after the α-layer low-frequency extraction at time t; S L is the total number of low-frequency decomposition layers; (u,v) is the pixel coordinate within the frame, satisfying 1≤u≤M, 1≤v≤N, and M and N are the width and height of the video frame respectively; for each low-frequency subband L α (t,u,v), calculate the energy difference between the current moment and the corresponding position at the previous moment (t-1), that is, D α (t,u,v) = |L α (t,u,v) - L α (t-1,u,v)|, where D α(t, u, v) represents the absolute difference of local illumination changes; perform a power transformation on the difference result, using the sensitivity parameter ρ (ρ > 0), and calculate where ρ is used to regulate the amplification effect on large changes; δ is an extremely small positive number; to obtain an overall illumination change index for one frame, perform an average normalization calculation on the transformation results of each low-frequency subband and all pixels to obtain the time-varying illumination fluctuation index.

[0016] In a preferred embodiment, the specific acquisition logic of the multi-scale noise stability index is as follows:

[0017] After the input video frame is decomposed by directionally controllable multi-scale decomposition, several high-frequency subband data are obtained, denoted as H β (t, u, v) (β = 1, 2, …, S H ); where H β (t, u, v) represents the pixel energy value after the β-th layer of high-frequency extraction at time t; S H is the total number of layers of high-frequency decomposition; to characterize the noise perturbation degree of each pixel within a single high-frequency subband, first calculate the global average value of each subband within the current frame: Subsequently, calculate the absolute deviation of each pixel position from the global mean: D β (t, u, v) = |H β (t, u, v) - μ β (t)|, perform an average normalization on each high-frequency subband within the full frame range to obtain the local noise stability coefficient of the corresponding subband; after the local noise stability coefficients of each high-frequency subband are weighted and aggregated, use logarithmic mapping to integrate the noise information of each scale, and finally obtain the multi-scale noise stability index NSI(t).

[0018] In a preferred embodiment, step S2 includes the following contents:

[0019] S2.1, select a sliding time window with a length of W, perform a Fourier transform on the time-varying illumination fluctuation index within each window, and extract the main frequency component with the largest amplitude, denoted as the main period of infrared supplementary light;

[0020] Within each window, calculate the theoretical fluctuation curve of the infrared supplementary light according to the main period of the infrared supplementary light, fit the corresponding periodic component using the least squares method, and subtract it from the time-varying illumination fluctuation index to obtain the compensated illumination fluctuation index;

[0021] Calculate the change rate of the illumination fluctuation index over time, and set an illumination mutation threshold. When the change of the illumination fluctuation index exceeds the illumination mutation threshold, the corresponding moment is considered an illumination mutation event and added to the illumination event sequence.

[0022] In a preferred embodiment, in S2.2, high-frequency analysis is performed on the input video frame using directionally controllable multi-scale decomposition, local peaks of each high-frequency sub-band are extracted, and the maximum energy values in different directions are recorded to form peak sets in multiple directions; specifically, for each high-frequency sub-band, a basic noise threshold is set and the dynamically adjusted threshold is obtained by dynamically correcting it using the multi-scale noise stability index, and the calculation is as follows:

[0023] where λ is an adjustment factor; at each moment t, check the peaks of all high-frequency sub-bands. If the peak in a certain direction exceeds the dynamically adjusted threshold θ noise,β (t), it is considered that the corresponding moment is a noise sudden increase event, and the corresponding moment is added to the noise event sequence.

[0024] In a preferred embodiment, in S2.3, since light mutation and noise sudden increase are related, it is necessary to calculate the matching degree between the two in time; first, define the time delay Δt = t i -t j , where t i is the light mutation moment, and t j is the noise sudden increase moment; use the time delay kernel function K(Δt) to perform weighted matching on all light mutation moments and noise sudden increase moments; calculate the total matching sum between all light mutation events and noise sudden increase events, and perform normalization adjustment using the multi-scale noise stability index. The final event correlation density AD is calculated as follows: Through non-linear mapping, the event correlation density is converted into the final global fusion selection coefficient, where |E light | represents the total number of light mutation events in the light event sequence; |E noise | represents the total number of noise sudden increase events in the noise event sequence; K(t i -t j ) is the time delay kernel function.

[0025] In a preferred embodiment, step S3 includes the following contents:

[0026] By comparing the global fusion selection coefficient with a preset critical value, it is determined as the high light-noise sensitive mode or the low light-noise stable mode according to the judgment, and the adjusted differential threshold is obtained; if the global fusion selection coefficient exceeds the preset critical value, activate the high light-noise sensitive mode, and amplify the basic differential threshold using an exponential response function, and the amplification ratio is proportional to the amplitude by which the global fusion selection coefficient exceeds the critical value; conversely, if the global fusion selection coefficient is lower than or equal to the preset critical value, enable the low light-noise stable mode, and perform attenuation adjustment on it by calculating the light variance of the current video frame using the light variance attenuation function.

[0027] In a preferred embodiment, step S4 includes the following contents:

[0028] Using the adjusted difference threshold D threshold Perform differential operation on the current video frame and the background model to extract the moving target area; first, calculate the pixel-level absolute difference between the current frame and the background model at each pixel position, and compare it with the adjusted differential threshold. If D(t,u,v)>D threshold , the corresponding pixel is marked as the moving target area, otherwise it is marked as the background, thereby generating a binary moving candidate area image;

[0029] Connected domain analysis is performed on the motion candidate area image to extract all connected areas and calculate the morphological features of each area. Then, all connected areas are screened using the preset water conservancy tool morphological template, and the morphological parameters of each connected area are compared with the water conservancy tool morphological template. If the morphological features of a certain area have a low matching degree with the template, the corresponding area is considered to be a noise artifact and is removed. Only the area that meets the tool morphological features is retained as the motion target candidate area set.

[0030] Multi-frame trajectory cross-validation is used to exclude transient false detection areas. First, the motion trajectory of the set of moving target candidate areas is tracked. For each moving target candidate area, its spatial position, area, and morphological changes are recorded in n consecutive frames, and the target motion trajectory is constructed. Then, the target motion trajectory is cross-validated to calculate the morphological change rate and position change rate of the corresponding area in adjacent frames. If the morphology of a certain area changes dramatically within multiple frames or the position drift exceeds the corresponding preset stability threshold, the corresponding area is determined to be transient noise or reflective pseudo-target and is eliminated. Only the area with stable morphology and motion is retained as the final moving target area.

[0031] In a preferred embodiment, step S5 includes the following contents:

[0032] The target motion trajectory is collected and cross-compared with the construction equipment vibration spectrum data and material transportation path data in terms of time series and spatial position to determine the correlation between the moving target and the construction machinery or material transportation behavior. When the correlation exceeds the set threshold, it means that some pixel fluctuations in the picture are not real targets, but background disturbances caused by mechanical vibration or material movement. In order to suppress the pixel degradation caused by these vibrations during the background model update process, a sensor-driven weight correction factor W is defined based on the comparison results. corr (t); Then, the weight correction factor is introduced into the update rate calculation of the background model, through the formula: α(t) = α base ×(1-λ1·W corr(t)); When updating the background model, a variable learning rate α(t) is used, where α base is the base update rate, and λ1 is an adjustment coefficient used to control the impact of the correction factor on the update rate.

[0033] An artificial intelligence-based water conservancy construction safety supervision system, including: a light noise analysis module, an event matching module, a mode switching module, a target screening module, and a background update module;

[0034] The light noise analysis module: performs frequency-domain sub-band energy analysis and directionally controllable multi-scale decomposition on real-time monitoring video frames, extracts the low-frequency energy change rate to generate a time-varying illumination fluctuation index respectively, and generates a multi-scale noise stability index through the high-frequency sub-band stability coefficient, and transfers the time-varying illumination fluctuation index and the multi-scale noise stability index to the event matching module;

[0035] The event matching module: performs time-delay kernel function matching on the illumination event sequence and the noise event sequence to calculate the event correlation density, and dynamically adjusts the noise determination threshold and matching normalization in combination with the multi-scale noise stability index, and finally generates a global fusion selection coefficient, and transfers the global fusion selection coefficient to the mode switching module;

[0036] The mode switching module: according to the comparison result between the global fusion selection coefficient and the critical value, respectively adjusts the base differential threshold using an exponential response or an illumination variance decay function to achieve dynamic switching between the high light noise sensitive mode and the low light noise stable mode, so as to optimize the moving target detection, and transfers the adjusted dynamic differential threshold to the target screening module;

[0037] The target screening module: performs a differential operation on the current frame and the background model using the adjusted dynamic differential threshold, screens the connected regions in combination with a preset morphological template of water conservancy tools, and eliminates instantaneous noise and reflective pseudo-targets through multi-frame trajectory cross-validation, and finally outputs the moving target area, and transfers the moving target area to the background update module;

[0038] The background update module: performs multi-source matching on the moving target trajectory, vibration spectrum, and transportation path, generates a sensor-driven weight correction factor according to the matching result, dynamically adjusts the background model update rate, and suppresses pixel degradation caused by mechanical vibration.

[0039] The technical effects and advantages of the artificial intelligence-based water conservancy construction safety supervision system and method of the present invention:

[0040] The present invention analyzes the energy of frequency sub-bands and performs directionally controllable multi-scale decomposition on real-time monitoring video frames, extracts the low-frequency energy change rate and the high-frequency sub-band stability coefficient respectively, generates a time-varying illumination fluctuation index and a multi-scale noise stability index, thereby accurately quantifying the dynamic characteristics of illumination and the level of noise interference, and effectively compensating for the periodic fluctuations of infrared supplementary lighting and environmental noise; subsequently, uses a time-delay kernel function to match the illumination events and noise events after dynamic compensation, and dynamically adjusts the noise determination threshold in combination with the noise stability index to generate a global fusion selection coefficient, realizing the efficient fusion of illumination mutation and noise sudden increase events; further, by comparing the global fusion selection coefficient with a preset critical value, adjusts the basic difference threshold using an exponential response function or an illumination variance attenuation function respectively, dynamically switches between a high-light-noise sensitive mode and a low-light-noise stable mode, and optimizes the detection accuracy of moving targets; performs a difference operation on the current frame and the background model using the adjusted difference threshold, screens the connected regions in combination with a preset morphological template of water conservancy tools, and eliminates instantaneous noise and reflection pseudo-targets through multi-frame trajectory cross-validation, and outputs an accurate moving target area; finally, performs multi-source matching on the moving target trajectory, the vibration spectrum of construction equipment, and the material transportation path data, generates a weight correction factor driven by sensors according to the matching result, dynamically adjusts the background model update rate, effectively suppresses pixel degradation caused by mechanical vibration, and constructs a closed-loop adaptive update mechanism, thereby significantly improving the real-time warning ability and safety supervision effect of the water conservancy construction site monitoring system. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 FIG. is a schematic flow chart of a water conservancy construction safety supervision method based on artificial intelligence according to the present invention;

[0042] Figure 2 FIG. is a schematic structural diagram of a water conservancy construction safety supervision system based on artificial intelligence according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0044] Example 1: Figure 1 A water conservancy construction safety supervision method based on artificial intelligence according to the present invention is provided, including:

[0045] S1: Perform frequency-domain sub-band energy analysis and directionally controllable multi-scale decomposition on the real-time monitoring video frames, respectively extract the low-frequency energy change rate to generate a time-varying illumination fluctuation index, and generate a multi-scale noise stability index through the high-frequency sub-band stability coefficient to quantify the illumination dynamic characteristics and noise interference level.

[0046] S2: Calculate the event correlation density by performing time-delay kernel function matching on the illumination event sequence and the noise event sequence after dynamic compensation, and dynamically adjust the noise determination threshold and matching normalization using the multi-scale noise stability index, and finally generate a global fusion selection coefficient.

[0047] S3: By comparing the global fusion selection coefficient with the critical value, respectively adjust the basic difference threshold using an exponential response or an illumination variance attenuation function to achieve dynamic switching between the high-light-noise sensitive mode and the low-light-noise stable mode to optimize the moving target detection.

[0048] S4: Perform a difference operation on the current frame and the background model using the adjusted difference threshold, screen the connected regions in combination with a preset morphological template of water conservancy tools, and eliminate instantaneous noise and reflective pseudo-targets through multi-frame trajectory cross-validation, and output the moving target region.

[0049] S5: Perform multi-source matching on the moving target trajectory, vibration spectrum, and transportation path, generate a sensor-driven weight correction factor according to the matching result, dynamically adjust the background model update rate, and suppress pixel degradation caused by mechanical vibration.

[0050] In the water conservancy construction environment, due to the fact that the illumination conditions are often affected by extreme meteorological factors such as sunlight transients, infrared supplementary lighting at night, and heavy rain and dust, resulting in drastic fluctuations, and at the same time, factors such as the vibration of construction equipment and local reflection will cause high-frequency noise interference. Therefore, it is necessary to accurately quantify the change trend of the low-frequency energy in the video frame by calculating the time-varying illumination fluctuation index to capture the dynamic changes and sudden changes of the overall illumination, providing an objective basis for subsequent dynamic compensation of the periodic fluctuations of infrared supplementary lighting and threshold regulation; while the multi-scale noise stability index uses the statistics of local noise deviation in the high-frequency sub-band to quantify the noise interference level caused by factors such as mechanical vibration, dust, and reflection, thereby effectively distinguishing real moving targets from artifact noise, ensuring the stable update of the background model and the accurate extraction of moving targets. The combined application of the two can achieve a comprehensive monitoring of the illumination and noise changes at the construction site, significantly improving the real-time warning accuracy of safety supervision.

[0051] Step S1 includes the following content:

[0052] Real-time monitored video frames may experience significant fluctuations due to sunlight transients, infrared light supplementation at night, and other external light interferences. To accurately capture the dynamic characteristics of light, it is necessary to perform directionally controllable multi-scale decomposition on each frame and extract the energy information reflecting the overall light trend in the low-frequency subband. By calculating the difference component of the low-frequency energy between the current frame and the previous frame, the light change rate can be quantified, and a time-varying light fluctuation index can be generated. This index will be used in subsequent steps to dynamically compensate the infrared light supplementation cycle and assist in determining the moment of sudden light change.

[0053] First, directionally controllable multi-scale decomposition is used to extract low-frequency information from the original video frames, focusing on capturing the global trends related to light. Subsequently, by calculating the absolute difference of the low-frequency energy between the current frame and the previous frame pixel by pixel, the local differences in light changes can be accurately reflected. After performing power transformation operations, both the impact of significant changes is enhanced, and the numerical noise caused by minor fluctuations is effectively suppressed. Finally, through global normalization and averaging, it is ensured that the generated time-varying light fluctuation index has stability and representativeness, and can be used to accurately locate the moment of sudden light change in subsequent steps, providing effective early warning information for construction safety monitoring.

[0054] The specific acquisition logic of the time-varying light fluctuation index is as follows:

[0055] The input video frame is denoted as F(t), with dimensions M×N. The input video frame is processed using a directionally controllable multi-scale decomposition operator to obtain several low-frequency subbands, denoted as: L α (t, u, v) (α = 1, 2, …, S L );

[0056] Where:

[0057] L α (t, u, v) represents the pixel value after low-frequency extraction at the α-th layer at time t;

[0058] S L is the total number of layers of low-frequency decomposition;

[0059] (u, v) are the pixel coordinates within the frame, satisfying 1 ≤ u ≤ M, 1 ≤ v ≤ N.

[0060] The directionally controllable filter kernel uses a pre-designed direction template in each decomposition layer to ensure that light information is fully retained during the extraction of each low-frequency subband while filtering out high-frequency noise components.

[0061] For each low-frequency subband L α (t, u, v), calculate the energy difference at the corresponding position between the current moment and the previous moment (t - 1), that is, D α (t, u, v) = |L α (t, u, v) - L α(t - 1, u, v)|, where D α (t, u, v) represents the absolute difference of local illumination change.

[0062] To enhance the sensitivity to large local illumination changes and reduce the influence of small noise, a power transformation is performed on the difference result. Using the sensitivity parameter ρ(ρ > 0), calculate

[0063] Where:

[0064] ρ is used to regulate the amplification effect on large changes, and the value is usually optimized according to experimental data;

[0065] δ is a very small positive number to prevent numerical instability when D α (t, u, v) is close to zero (e.g., δ = 10 -6 ).

[0066] To obtain an overall illumination change index for one frame, the transformation results on each low-frequency subband and all pixels are averaged and normalized to obtain the time-varying illumination fluctuation index TIF(t):

[0067] Here, the summation and normalization operations ensure that the index reflects the global illumination change trend rather than the influence of local noise or outliers.

[0068] First, the noise fluctuation index is obtained by calculating the local energy deviation within each high-frequency subband, and then the stability coefficient is obtained by averaging and normalizing the entire frame of pixels. Based on the weight parameters determined by historical data statistics or engineering experience, the noise coefficients at each scale are weighted and integrated, and finally, a logarithmic mapping is used to generate the multi-scale noise stability index. This calculation process ensures that in a complex construction environment, the high-frequency noise interference level can be fully quantified, providing an accurate and stable index basis for subsequent dynamic threshold regulation and noise mutation detection.

[0069] The specific acquisition logic of the multi-scale noise stability index is as follows:

[0070] After the input video frame is decomposed by directionally controllable multi-scale decomposition, several high-frequency subband data are obtained, denoted as H β (t, u, v) (β = 1, 2,..., S H );

[0071] Where:

[0072] H β (t, u, v) represents the pixel energy value after the β-th layer of high-frequency extraction at time t;

[0073] S His the total number of layers for high-frequency decomposition;

[0074] (u, v) are pixel coordinates, satisfying 1 ≤ u ≤ M and 1 ≤ v ≤ N, where M and N are the width and height of the video frame respectively.

[0075] To characterize the noise perturbation degree of each pixel within a single high-frequency subband, first calculate the global average value of each subband within the current frame:

[0076] Subsequently, calculate the absolute deviation of each pixel position from the global mean: D β (t, u, v) = H β (t, u, v) - μ β (t)|, and the absolute deviation reflects the intensity of local noise fluctuations.

[0077] To reduce the interference of pixel-level accidental fluctuations, perform average normalization on each high-frequency subband within the entire frame range to obtain the local noise stability coefficient S β (t):

[0078] Here, S β (t) represents the average noise fluctuation level of the β-th high-frequency subband at time t.

[0079] Among high-frequency subbands with different directions and scales, the sensitivity to noise information may vary. To reflect this difference, introduce the weight parameter ω β to regulate the contribution of each subband.

[0080] Explanation of the weight source:

[0081] The weight ω β can be obtained through statistical analysis of historical monitoring data. For example, in long-term monitoring data, calculate the energy standard deviation σ β or the noise contribution ratio for each high-frequency subband, and then through normalization processing to make:

[0082] Or determine it based on on-site experiments and engineering experience, so that subbands with higher sensitivity obtain larger weights. This method ensures that each weight satisfies the normalization condition, that is:

[0083] The weight parameter makes the contribution of each subband to the final noise stability index more in line with the actual noise interference characteristics.

[0084] After weighted aggregation of the local noise stability coefficients of each high-frequency subband, to enhance the response to large fluctuations and compress the numerical range, use logarithmic mapping to integrate the noise information at each scale. Finally, define the multi-scale noise stability index as NSI(t):

[0085]

[0086] Wherein:

[0087] The function of the logarithmic function is to smooth large - range changes and make the exponential values more stable;

[0088] Adding 1 ensures that the exponent is still defined when the noise fluctuation is extremely small;

[0089] The weighted sum of weights ensures that the noise information of each high - frequency sub - band is reasonably integrated according to the sensitivity.

[0090] The multi - scale noise stability index describes the comprehensive level of high - frequency noise interference in the current frame. The larger the value, the more significant the noise interference.

[0091] The dynamic and uncertain nature of the water conservancy construction environment makes video surveillance face complex illumination changes and high - frequency noise interference. Especially during night construction, the periodic fluctuations of infrared supplementary light are intertwined with factors such as the vibration of construction equipment and dust reflection, making it difficult for traditional moving target detection methods to distinguish real changes from artifact interference. When the periodic change of infrared supplementary light is not compensated, sudden illumination changes are easily misjudged as target motion, and if the sudden increase in high - frequency noise lacks accurate identification, it may lead to chaos in background model update, further exacerbating the problems of false detection and missed detection. Therefore, a technical solution that can simultaneously compensate for illumination fluctuations, screen sudden increases in noise, and analyze the correlation between the two types of events in the time dimension is required to ensure the accuracy of moving target detection. The goal of step S2 is to extract illumination mutation information using the time - varying illumination fluctuation index, dynamically screen the moments of sudden increase in noise in combination with the multi - scale noise stability index, and calculate the time matching degree of the two types of events through a time - delay kernel function to generate a global fusion selection coefficient, providing reliable data support for subsequent dynamic threshold regulation and background model adaptive update.

[0092] Step S2 includes the following contents:

[0093] S2.1, Since infrared supplementary light usually changes periodically, it is necessary to extract the dominant frequency information in the time - varying illumination fluctuation index. First, a sliding time window with a length of W is selected, and the Fourier transform of the time - varying illumination fluctuation index is performed within each window to extract the main frequency component with the largest amplitude, denoted as the main period of infrared supplementary light.

[0094] Within each window, according to the main period of infrared supplementary light, calculate the theoretical fluctuation curve of infrared supplementary light, fit the corresponding periodic component using the least - squares method, and subtract it from the time - varying illumination fluctuation index to obtain the compensated illumination fluctuation index. This can remove the periodic influence of infrared supplementary light, making the remaining illumination fluctuation part better reflect the real illumination change.

[0095] Calculate the change rate of the light intensity fluctuation index over time, and set a threshold for light intensity mutation. When the change of the light intensity fluctuation index exceeds the threshold for light intensity mutation, the moment is considered a light intensity mutation event and is added to the light intensity event sequence.

[0096] S2.2. Perform high-frequency analysis on the input video frame using directionally controllable multi-scale decomposition, extract the local peaks of each high-frequency sub-band, and record the maximum energy values in different directions to form peak sets in multiple directions.

[0097] Since the noise level is dynamically changing, a fixed threshold cannot be used to screen for sudden increases in noise. Therefore, the multi-scale noise stability index is used to adjust the noise detection threshold for each sub-band, so that the threshold is relaxed when the noise level is high to avoid excessive false alarms, and the threshold is tightened when the noise level is low to prevent missed detections.

[0098] Specifically, for each high-frequency sub-band, a basic noise threshold is set and dynamically corrected using the multi-scale noise stability index to obtain the dynamically adjusted threshold, which is calculated as follows:

[0099] where λ is a regulation factor used to control the influence of the noise index on the threshold adjustment.

[0100] At each moment t, check the peaks of all high-frequency sub-bands. If the peak in a certain direction exceeds the dynamically adjusted threshold θ noise,β (t), then the moment is considered a sudden increase in noise event and the moment is added to the noise event sequence.

[0101] S2.3. Since light intensity mutation and sudden increase in noise may be related, it is necessary to calculate the matching degree between the two in time. First, define the time delay Δt = t i -t j , where t i is the moment of light intensity mutation and t j is the moment of sudden increase in noise.

[0102] Use the time-delay kernel function K(Δt) to perform weighted matching on all possible moments of light intensity mutation and sudden increase in noise. The role of the kernel function is to give higher matching weights to events with closer times and lower weights to events with larger time differences. The kernel function is calculated as follows: where σ t is the sensitive scale of time-delay matching used to control the matching accuracy.

[0103] Calculate the total match between all light mutation events and noise spike events, and perform normalization adjustment using the multi-scale noise stability index to reduce the sensitivity of the matching result in the case of high noise and prevent the false high event matching degree caused by noise interference. The final event association density AD is calculated as follows:

[0104] Through non-linear mapping, convert the event association density into the final global fusion selection coefficient GFC: GFC = 1 - exp(-AD);

[0105] Where:

[0106] |E light | represents the total number of events in the light event sequence;

[0107] |E noise | represents the total number of events in the noise event sequence;

[0108] K(t i -t j ) is the time delay kernel function.

[0109] The global fusion selection coefficient reflects the time correlation between light mutations and noise spike events. The larger the value, the closer the relationship between the two, providing a basis for subsequent dynamic threshold adjustment and background model update.

[0110] Step S3 includes the following:

[0111] By comparing the global fusion selection coefficient with a preset critical value, clarify the light and noise interference environment of the current video frame, and determine the high light-noise sensitive mode or low light-noise stable mode according to the judgment, and obtain the adjusted difference threshold. If the global fusion selection coefficient exceeds the preset critical value, it indicates that there are drastic light changes and noise interference in the scene. At this time, activate the high light-noise sensitive mode; in this mode, use the exponential response function to amplify the basic difference threshold, and the amplification ratio is proportional to the amplitude by which the global fusion selection coefficient exceeds the critical value, so as to enhance the sensitivity of the background difference operation to the mutation signal, so that the light changes caused by infrared supplementary light or sudden light changes can be quickly reflected in the difference result, facilitating the capture of real target movements; on the contrary, if the global fusion selection coefficient is lower than or equal to the preset critical value, it is determined that the current environment is relatively stable. At this time, enable the low light-noise stable mode, calculate the light variance of the current video frame, and use the light variance attenuation function to perform attenuation adjustment, so that the threshold is reduced under the condition of low noise interference, thereby suppressing false detections caused by weak noise; the adjusted difference threshold, whether it is amplified exponentially or attenuated by light variance, will be directly used for subsequent background model update and moving target detection to ensure that in different light and noise environments, real moving targets can be accurately distinguished from environmental noise and dynamic adaptation can be achieved.

[0112] Moving targets in surveillance videos not only include real construction equipment, personnel, and material movement, but may also be interfered by factors such as infrared supplementary lighting changes, mechanical vibrations, and dust reflections, resulting in false detections and missed detections easily occurring in the extraction of moving targets by traditional background modeling methods. Especially at night or under extreme weather conditions, sudden changes in light and high-frequency noise interference may cause a large number of pseudo-targets to appear in the background difference results, and traditional morphological filtering or single-frame difference methods are difficult to effectively remove these short-term fluctuation noises. Therefore, it is necessary to perform background difference operation based on the dynamically adjusted threshold after step S3 to ensure that the moving target detection has self-adaptability to light fluctuations and noise interference. At the same time, combine the morphological template of water conservancy tools to screen the connected regions that conform to the tool characteristics, and analyze the morphology and movement consistency of the moving targets in consecutive frames through multi-frame trajectory cross-verification, remove the noise regions with short-term fluctuations and reflective pseudo-targets, and finally achieve accurate extraction of moving targets, providing stable and reliable data support for subsequent safety supervision of the construction site.

[0113] Step S4 includes the following content:

[0114] Use the adjusted difference threshold D output by step S3 threshold Perform a difference operation on the current video frame F(t) and the background model B(t) to extract the possible moving target regions. First, calculate the pixel-level absolute difference D(t,u,v) = |F(t,u,v) - B(t,u,v)| between the current frame and the background model at each pixel position, and compare it with the adjusted difference threshold. If D(t,u,v) > D threshold , then mark this pixel as a moving target region, otherwise mark it as the background, thus generating a binary moving candidate region image. Since the direct difference result may contain noise interference or non-real targets, it is necessary to further optimize the extraction of moving targets.

[0115] First, perform connected component analysis on the moving candidate region image, extract all connected regions, and calculate the morphological characteristics of each region, including area, aspect ratio, contour complexity, etc. Then, use the preset morphological template of water conservancy tools to screen all connected regions. Compare the morphological parameters of each connected region with the morphological template of water conservancy tools. If the morphological characteristics of a certain region have a low matching degree with the template, it is considered that this region may be noise artifacts and should be excluded, and only the regions that conform to the morphological characteristics of the tools are retained as the set of moving target candidate regions.

[0116] Since single-frame detection is easily interfered by instantaneous noise and light reflection, multi-frame trajectory cross-verification is further used to exclude transient false detection regions. First, perform motion trajectory tracking on the set of moving target candidate regions. For each moving target candidate region R i(t), record its spatial position, area, and morphological changes within consecutive n frames, and construct the target motion trajectory where (x k , y k ) represents the central position of the region, and A k represents its area. Then, perform cross-validation on the target motion trajectory, and calculate the morphological change rate ΔA k = |A k - A k-1 | / A k-1 and the position change rate If a certain region has a drastic morphological change or a position drift exceeding the corresponding preset stability threshold within multiple frames, then determine that this region is transient noise or a reflective pseudo-target, and eliminate it, only retaining the regions with stable morphology and motion as the final motion target regions.

[0117] Step S5 includes the following content:

[0118] First, collect the target motion trajectories obtained in the previous step S4, and perform cross-comparisons in terms of time series and spatial positions with the vibration spectrum data of construction equipment and the material transportation path data to determine the correlation degree between the moving target and the construction machinery or material transportation behavior. For example, record the coordinate points passed by a certain target within a certain time period. At the same time, obtain the vibration spectrum data of the construction equipment, which reflects the vibration energy distribution and main frequency components of the equipment at different times, and the material transportation path data, which records the actual movement routes and position changes of transportation vehicles or equipment within the construction site. Next, compare the time period of the target motion trajectory with the vibration data to check whether obvious frequency peaks related to the operation of the construction equipment are detected in the time interval where the target passes. For example, when a certain target appears near the construction equipment within a certain time period, and at the same time, significant peaks of the equipment characteristic frequencies are shown in the vibration data during this time period, it can be considered that there is a temporal correlation between the target motion and the construction equipment vibration. At the same time, cross-compare the spatial position of the target motion trajectory with the material transportation path to determine whether the path passed by the target coincides with or is close to the running route of the transportation vehicle. For example, if the distance between the area where the target is located and the area passed by the transportation vehicle is less than a predetermined distance, it indicates that there is an obvious correlation between the two in space. Finally, weight and integrate the results of time matching and spatial matching according to the preset weights to calculate a correlation degree. When the correlation degree exceeds the set threshold, it can be confirmed that there is a high correlation between the moving target and the construction equipment or material transportation behavior, thereby providing a data basis for the generation of the subsequent sensor-driven weight correction factor. When the correlation degree exceeds the set threshold, it means that some pixel fluctuations in the picture may not be real targets, but background disturbances caused by mechanical vibrations or material movements. In order to suppress the pixel degradation caused by these vibrations during the background model update process, a sensor-driven weight correction factor W corr (t) is defined according to the comparison results. The larger the value, the more significant the impact of mechanical vibrations or material disturbances on the current frame. Then, introduce the weight correction factor into the calculation of the update rate of the background model, through the formula: α(t) = α base ×(1 - λ1·W corr (t));

[0119] Use a variable learning rate α(t) when updating the background model B(t), where α baseFor the base update rate, λ1 is an adjustment coefficient used to control the influence of the correction factor on the update rate; when the weight correction factor is large, it indicates frequent mechanical vibrations or material transportation interferences, and the background update rate needs to be reduced to prevent the normal background from being quickly replaced. Conversely, it gradually returns to a higher update rate to adapt to the gradual environmental changes and reduce cumulative errors. Through the fusion of such sensors and video information, a closed-loop adaptive update mechanism is constructed, which can accurately suppress mechanical vibrations and transportation interferences in real-time monitoring, ensuring the long-term stability of the background model for dynamic construction scenes.

[0120] The vibration spectrum data of construction equipment refers to the distribution information obtained by performing frequency-domain analysis on the vibration signals generated during the operation of mechanical equipment. Usually, accelerometers or other types of vibration sensors are installed at key parts of the equipment to record the energy amplitudes of different frequency components, thereby reflecting the operating conditions and vibration characteristics of the equipment; the material transportation path data refers to the continuous recording of the travel trajectories of material transportation vehicles, loading equipment, etc. at the construction site in space and time. Usually, information such as the driving route, stopping position, and time distribution is obtained through GPS or other positioning means to help determine whether the movement appearing in the picture is related to the transportation operation; the weight correction factor is a dynamic coefficient calculated based on the matching degree between the moving target and the above vibration spectrum or transportation path, and is used to adjust the learning rate during the update of the background model. When a pixel change is identified as a disturbance caused by mechanical vibration or material transportation, the correction factor is increased to reduce the possibility of rapid background replacement, thereby avoiding misjudging the real background as a moving target or causing pixel degradation problems.

[0121] Embodiment 2: Figure 2 A water conservancy construction safety supervision system based on artificial intelligence according to the present invention is given, including: a light noise analysis module, an event matching module, a mode switching module, a target screening module, and a background update module;

[0122] The light noise analysis module: performs frequency-domain sub-band energy analysis and directionally controllable multi-scale decomposition on the real-time monitoring video frames, extracts the low-frequency energy change rate to generate a time-varying illumination fluctuation index respectively, and generates a multi-scale noise stability index through the high-frequency sub-band stability coefficient, and transmits the time-varying illumination fluctuation index and the multi-scale noise stability index to the event matching module;

[0123] The event matching module: performs time-delay kernel function matching on the illumination event sequence and the noise event sequence to calculate the event correlation density, and dynamically adjusts the noise determination threshold and matching normalization in combination with the multi-scale noise stability index, and finally generates a global fusion selection coefficient, and transmits the global fusion selection coefficient to the mode switching module;

[0124] Mode switching module: According to the comparison result between the global fusion selection coefficient and the critical value, the exponential response or the light variance attenuation function is used to adjust the basic differential threshold respectively, so as to realize the dynamic switching between the high-light-noise sensitive mode and the low-light-noise stable mode, optimize the moving target detection, and transfer the adjusted dynamic differential threshold to the target screening module;

[0125] Target screening module: Use the adjusted dynamic differential threshold to perform differential operation on the current frame and the background model, screen the connected regions in combination with the preset morphological template of water conservancy tools, and eliminate instantaneous noise and reflective pseudo-targets through multi-frame trajectory cross-verification. Finally, output the moving target area and transfer the moving target area to the background update module;

[0126] Background update module: Perform multi-source matching on the moving target trajectory, vibration spectrum and transportation path, generate a sensor-driven weight correction factor according to the matching result, dynamically adjust the background model update rate, and suppress pixel degradation caused by mechanical vibration.

[0127] The above formulas are all dimensionless and take their numerical calculations. The formulas are obtained by collecting a large amount of data for software simulation to get a formula closest to the actual situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0128] Only some exemplary embodiments of the present invention are described above by way of illustration. Undoubtedly, for those of ordinary skill in the art, without departing from the spirit and scope of the present invention, the described embodiments can be modified in various different ways. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the protection scope of the claims of the present invention.

[0129] It should be noted that in this article, if there are relational terms such as first and second, they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of another identical element in the process, method, article or device including the element.

[0130] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claimed rights.

Claims

1. A method for safety supervision of water conservancy construction based on artificial intelligence, characterized in that, Including the steps: S1: Perform frequency-domain sub-band energy analysis and directionally controllable multi-scale decomposition on the real-time monitoring video frames, extract the low-frequency energy change rate to generate the time-varying illumination fluctuation index respectively, and generate the multi-scale noise stability index through the high-frequency sub-band stability coefficient to quantify the illumination dynamic characteristics and noise interference level; S2: Calculate the event correlation density by performing time-delay kernel function matching on the illumination event sequence and noise event sequence after dynamic compensation, and use the multi-scale noise stability index to dynamically adjust the noise determination threshold and matching normalization, and finally generate the global fusion selection coefficient; S3: By comparing the global fusion selection coefficient with the critical value, adjust the basic difference threshold respectively using the exponential response or illumination variance attenuation function to achieve dynamic switching between the high-light-noise sensitive mode and the low-light-noise stable mode to optimize the moving target detection; S4: Perform difference operation on the current frame and the background model using the adjusted difference threshold, screen the connected regions in combination with the preset morphological template of water conservancy tools, and eliminate instantaneous noise and reflective pseudo-targets through multi-frame trajectory cross-verification, and output the moving target region; S5: Perform multi-source matching on the moving target trajectory, vibration spectrum and transportation path, generate a sensor-driven weight correction factor according to the matching result, and dynamically adjust the background model update rate and suppress pixel degradation caused by mechanical vibration.

2. The water conservancy construction safety supervision method based on artificial intelligence according to claim 1, wherein, Step S1 includes the following contents: The specific acquisition logic of the time-varying illumination fluctuation index is: The input video frame is denoted as F(t), whose size is M×N. The input video frame is processed using a steerable multi-scale decomposition operator to obtain several low-frequency sub-bands, denoted as: L α (t,u,v)(α=1,2,…,S L );where L α (t,u,v) represents the pixel value after the αth layer low-frequency extraction at time t; S L is the total number of low-frequency decomposition layers; (u, v) is the pixel coordinate within the frame, satisfying 1≤u≤M, 1≤v≤N, where M and N are the width and height of the video frame respectively; for each low-frequency subband L α (t,u,v), calculate the energy difference between the current moment and the corresponding position at the previous moment (t-1), that is, D α (t,u,v)=|L α (t,u,v)-L α (t-1,u,v)|, where D α (t,u,v) represents the absolute difference of local illumination change; the difference result is transformed to power, and the sensitivity parameter ρ (ρ>0) is used to calculate Where ρ is used to control the amplification effect of large changes; δ is a very small positive number; to obtain the overall illumination change index of a frame, the transformation results of each low-frequency sub-band and all pixels are averaged and normalized to obtain the time-varying illumination fluctuation index.

3. The method for safety supervision of water conservancy construction based on artificial intelligence according to claim 2, characterized in that, The specific acquisition logic of the multi-scale noise stability index is: After the input video frame is decomposed by directionally controllable multi-scale decomposition, several high-frequency sub-band data are obtained, denoted as H β (t, u, v) (β = 1, 2, …, S H ); where H β (t, u, v) represents the pixel energy value after high-frequency extraction at the β-th layer at time t; SH is the total number of high-frequency decomposition layers; to characterize the noise perturbation degree of each pixel in a single high-frequency sub-band, first calculate the global average value of each sub-band in the current frame: Subsequently, calculate the absolute deviation of each pixel position from the global mean: D β (t, u, v) = |H β (t, u, v) - μ β (t)|, and perform average normalization on each high-frequency sub-band within the full frame range to obtain the local noise stability coefficient of the corresponding sub-band. After the local noise stability coefficients of each high-frequency sub-band are weighted and summarized, the logarithmic mapping is used to integrate the noise information of each scale, and finally the multi-scale noise stability index NSI(t) is obtained.

4. The method for safety supervision of water conservancy construction based on artificial intelligence according to claim 3, characterized in that, Step S2 includes the following contents: S2.1, select a sliding time window with a length of W, perform Fourier transform on the time-varying illumination fluctuation index within each window, and extract the main frequency component with the largest amplitude, denoted as the main period of infrared supplementary light; Within each window, calculate the theoretical fluctuation curve of the infrared supplementary light according to the main period of the infrared supplementary light, fit the corresponding periodic component using the least squares method, and subtract it from the time-varying illumination fluctuation index to obtain the compensated illumination fluctuation index; Calculate the change rate of the illumination fluctuation index in time, and set the illumination mutation threshold. When the change of the illumination fluctuation index exceeds the illumination mutation threshold, the corresponding moment is considered as an illumination mutation event and added to the illumination event sequence.

5. A water conservancy construction safety supervision method based on artificial intelligence according to claim 4, characterized in that: S2.

2. Perform high-frequency analysis on the input video frames using directionally controllable multi-scale decomposition, extract the local peaks of each high-frequency subband, and record the maximum energy values in different directions to form peak sets in multiple directions. Specifically, for each high-frequency subband, set a basic noise threshold and obtain the dynamically adjusted threshold by dynamically correcting it with the multi-scale noise stability index, which is calculated as follows: where λ is a regulation factor; at each moment t, the peaks of all high-frequency subbands are checked, and if the peak in a certain direction exceeds the dynamically adjusted threshold θ noise,β ( t ) , it is considered that the corresponding moment is a noise surge event, and the corresponding moment is added to the noise event sequence.

6. A water conservancy construction safety supervision method based on artificial intelligence according to claim 5, characterized in that: S2.

3. Since the light mutation and the sudden increase in noise are related, it is necessary to calculate the degree of their temporal matching. First, define the time delay Δt = t i -t j , where t i is the moment of light mutation, and t j is the moment of sudden increase in noise. Use the time-delay kernel function K ( Δt ) to perform weighted matching for all moments of light mutation and sudden increase in noise. Calculate the total matching between all light mutation events and sudden increase in noise events, and use the multi-scale noise stability index for normalization adjustment. The final event correlation density AD is calculated as follows: Through non-linear mapping, convert the event correlation density into the final global fusion selection coefficient, where | E light | represents the total number of light mutation events in the light event sequence; | E noise| represents the total number of sudden increase in noise events in the noise event sequence; K ( t i -t j ) is the time-delay kernel function.

7. The method for safety supervision of water conservancy construction based on artificial intelligence according to claim 6, wherein, Step S3 includes the following contents: By comparing the global fusion selection coefficient with the preset critical value, determine whether it is in the high-light-noise sensitive mode or the low-light-noise stable mode according to the judgment, and obtain the adjusted difference threshold; If the global fusion selection coefficient exceeds the preset critical value, activate the high-light-noise sensitive mode, and amplify the basic difference threshold using the exponential response function, and the amplification ratio is proportional to the amplitude by which the global fusion selection coefficient exceeds the critical value; On the contrary, if the global fusion selection coefficient is lower than or equal to the preset critical value, the low light noise stabilization mode is enabled, and the illumination variance of the current video frame is calculated and the illumination variance attenuation function is used to perform attenuation adjustment.

8. The method for safety supervision of water conservancy construction based on artificial intelligence according to claim 7, characterized in that, Step S4 includes the following contents: Using the adjusted difference threshold D threshold Perform a difference operation on the current video frame and the background model to extract the moving target area; First, calculate the pixel-level absolute difference between the current frame and the background model at each pixel position, and compare it with the adjusted difference threshold. If D ( t,u,v ) >D threshold , mark the corresponding pixel as the moving target area, otherwise mark it as the background, thereby generating a binary moving candidate area image; Conduct connected domain analysis on the motion candidate region image, extract all connected regions, and calculate the morphological features of each region; Then, all connected regions are screened using the preset morphological template of hydraulic tools. The morphological parameters of each connected region are compared with the morphological template of hydraulic tools. If the morphological features of a region have a low matching degree with the template, the corresponding region is considered to be a noise artifact and is removed. Only the region that meets the morphological features of tools is retained as the candidate region set of moving targets. Multi-frame trajectory cross-validation is used to exclude transient false detection areas. First, the motion trajectory of the set of moving target candidate areas is tracked. For each moving target candidate area, its spatial position, area, and morphological changes are recorded in n consecutive frames, and the target motion trajectory is constructed. Then, the target motion trajectory is cross-validated to calculate the morphological change rate and position change rate of the corresponding area in adjacent frames. If the morphology of a certain area changes dramatically within multiple frames or the position drift exceeds the corresponding preset stability threshold, the corresponding area is determined to be transient noise or reflective pseudo-target and is eliminated. Only the area with stable morphology and motion is retained as the final moving target area.

9. The water conservancy construction safety supervision method based on artificial intelligence according to claim 8, characterized in that, Step S5 includes the following contents: Collect the obtained target motion trajectory and perform cross-comparisons in terms of time series and spatial position with the vibration spectrum data of the construction equipment and the material transportation path data to determine the correlation between the moving target and the construction machinery or material transportation behavior. When the correlation exceeds the set threshold, it indicates that some pixel fluctuations in the picture are not real targets but background disturbances caused by mechanical vibrations or material movements. In order to suppress the pixel degradation caused by these vibrations during the background model update process, a sensor-driven weight correction factor W corr (t); Then, a weight correction factor is introduced into the calculation of the update rate of the background model, through the formula: α(t) = α base ×(1 - λ1·W corr (t)); A variable learning rate α(t) is used when updating the background model, where α base is the basic update rate, and λ1 is a regulation coefficient used to control the influence of the correction factor on the update rate.

10. A water conservancy construction safety supervision system based on artificial intelligence, which is used to implement a water conservancy construction safety supervision method according to any one of claims 1-9, and is characterized in that, include: Light noise analysis module, event matching module, mode switching module, target screening module and background update module; Optical noise analysis module: performs frequency domain sub-band energy analysis and direction-controllable multi-scale decomposition on real-time surveillance video frames, extracts low-frequency energy change rates to generate time-varying illumination fluctuation indexes, and generates multi-scale noise stability indexes through high-frequency sub-band stability coefficients, and passes the time-varying illumination fluctuation indexes and multi-scale noise stability indexes to the event matching module; Event matching module: performs time delay kernel function matching on the illumination event sequence and the noise event sequence to calculate the event correlation density, and dynamically adjusts the noise judgment threshold and matching normalization in combination with the multi-scale noise stability index, and finally generates a global fusion selection coefficient, which is passed to the mode switching module; Mode switching module: According to the comparison result of the global fusion selection coefficient and the critical value, the basic differential threshold is adjusted by using the exponential response or the illumination variance attenuation function to realize the dynamic switching between the high light noise sensitive mode and the low light noise stable mode to optimize the detection of moving targets, and the adjusted dynamic differential threshold is passed to the target screening module; Target screening module: Use the adjusted dynamic differential threshold to perform differential operation on the current frame and the background model, combine the preset morphological template of water conservancy tools to screen the connected domain, and eliminate instantaneous noise and reflective pseudo targets through multi-frame trajectory cross-validation, and finally output the moving target area, and pass the moving target area to the background update module; Background update module: Perform multi-source matching of the moving target trajectory, vibration spectrum, and transportation path, generate a sensor-driven weight correction factor based on the matching result, dynamically adjust the background model update rate, and suppress pixel degradation caused by mechanical vibration.

Images