AI-based intelligent security central control event analysis and decision support system

Abstract

The application discloses an AI-based intelligent security and protection central control event analysis and decision support system, relates to the technical field of intelligent security and protection and computer vision, and comprises a shielding recognition module configured to receive a monitoring video stream, recognize dynamic shielding objects and their covered areas in a video picture, and the dynamic shielding object is a non-rigid visual shielding medium which evolves dynamically with time; and a motion feature extraction module configured to extract motion feature information of a steam cloud edge area, combine a long short-term memory network to learn a nonlinear mapping relationship between steam cloud edge optical flow and internal target motion, realize motion state inference from a visible edge to an invisible target, successfully solve the problem that a target cannot be recognized and tracking is interrupted under steam cloud shielding, and greatly improve the continuity of monitoring.

Description

Technical Field

[0001] This invention relates to the fields of intelligent security and computer vision technology, specifically to an AI-based intelligent security central control event analysis and decision support system. Background Technology

[0002] In industrial settings such as chemical industrial parks, high-pressure steam generated by safety valve depressurization and process venting, along with low-temperature white fog formed by leaks of cryogenic media such as liquefied natural gas, together constitute steam clouds characterized by suddenness, high density, and dynamic evolution. When a steam cloud envelops a surveillance camera, the visible light image appears completely white or gray, causing the edges and textures of targets such as people, vehicles, and open flames to be completely obscured, resulting in an image signal-to-noise ratio approaching 0dB. This is the primary technical challenge in chemical safety scenarios.

[0003] Traditional AI vision models rely on edge features and gradient operators to extract target information. In extreme visual environments caused by vapor cloud obstruction, they cannot extract any effective information, thus falling into the dilemma of "sensor blindness"—they cannot determine "no event" (to avoid missing potential dangerous leaks) nor can they determine "camera malfunction" and ignore monitoring. This ultimately leads to a large number of false alarms and missed alarms, which seriously weakens the reliability of intelligent security systems and cannot meet the safety monitoring needs of chemical industry scenarios.

[0004] Existing solutions to the above problems all have limitations and fail to resolve the core contradictions: First, multimodal sensor fusion (such as the combination of visible light and infrared thermal imaging) not only has limited distinguishability in low-temperature leakage scenarios, but also suffers from high costs, making it difficult to promote and apply on a large scale; Second, image completion and frame interpolation techniques assume that the obstruction is static and repairable, which cannot adapt to the non-rigid and rapidly evolving characteristics of steam clouds, resulting in poor repair effects; Third, temporal memory tracking technology relies too heavily on the complete target features before obstruction, making it extremely difficult to re-identify the target after prolonged obstruction by steam clouds, thus failing to achieve continuous target tracking.

[0005] The core technical challenge lies in the fact that vapor clouds are both obstructions that cause visual monitoring to fail and key signals that may indicate dangerous leaks. Existing technologies have failed to combine the two, making it impossible to infer the internal target state by using the movement characteristics of the obstruction itself when the obstructed target is completely invisible, and also impossible to combine chemical process data to determine the nature of the vapor cloud (normal emission or abnormal leakage). This has become a core technical challenge that chemical safety control systems urgently need to overcome. Summary of the Invention

[0006] The purpose of this invention is to provide an AI-based intelligent security central control event analysis and decision support system to address the shortcomings of the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an AI-based intelligent security central control event analysis and decision support system, comprising: The occlusion recognition module is configured to receive the monitoring video stream and identify dynamic occlusion objects and their coverage areas in the video frame. The dynamic occlusion objects are non-rigid visual occlusion media that evolve dynamically over time. The motion feature extraction module is configured to extract motion feature information of the edge region of the dynamic occlusion when the target features in the video image cannot be normally identified due to the dynamic occlusion. The motion feature information refers to the optical flow vector field of the pixels on the edge contour of the steam cloud, where each optical flow vector contains pixel position coordinates, velocity vector, orientation angle, and amplitude information. The steam cloud edge region is the set of pixels determined by extracting its contour after identifying the steam cloud coverage area using a deep semantic segmentation network. The targets include personnel, equipment, vehicles, and open flames in a chemical scene. The state inference module is configured to infer the motion state of the occluded target during the occlusion period based on the motion feature information, and generate a complete trajectory during the occlusion period. The identity inheritance module is configured to match the completed trajectory with the restored target after the occlusion dissipates, thereby achieving target identity inheritance.

[0008] In a preferred embodiment, the dynamic occlusion is specifically a steam cloud in a chemical industry setting, and the motion feature extraction module includes: The vapor cloud edge detection unit is configured to use a deep semantic segmentation network to identify the vapor cloud coverage area and extract the edge contour; The optical flow calculation unit is configured to use an improved Lucas-Kanade optical flow method based on an image pyramid structure to calculate the optical flow vector field of the pixels at the edge contour of the vapor cloud. The optical flow feature construction unit is configured to construct an optical flow feature matrix, where each optical flow feature contains pixel position coordinates, velocity vector, orientation angle, and amplitude information; The target detection unit is configured to perform target detection on a visible light image under unobstructed or weakly obstructed conditions and output the target motion trajectory. The weak occlusion condition is defined as follows: the proportion of the vapor cloud coverage area is less than the first weak occlusion threshold, and the target detection confidence is greater than the second weak occlusion threshold; the value range of the first weak occlusion threshold is 10% to 20%, and the value range of the second weak occlusion threshold is 0.6 to 0.8.

[0009] In a preferred embodiment, the motion feature extraction module further includes an optical flow noise suppression unit, configured as follows: The vapor cloud region is divided into a core region and an edge transition region. The core region is the pixel value saturation area, and the edge transition region is the gradient change zone extending outward from the core region. The optical flow in the core region is marked as a noise source, and bidirectional dynamic threshold segmentation is used to remove optical flow noise in the core region. Median filtering is applied to the optical flow data retained in the edge transition region to remove isolated noise points.

[0010] In a preferred embodiment, the optical flow noise suppression unit is configured to perform the following when executing the bidirectional dynamic thresholding: Horizontal segmentation: Threshold segmentation is performed on the optical flow component u in the x-direction, with the threshold value... ,in The mean of the u-component. The standard deviation is denoted by α, which is an adjustable coefficient ranging from 1.5 to 3.0. Vertical segmentation: Threshold segmentation is performed on the optical flow component v in the y-direction, with the threshold value... ,in The mean of the v component. The standard deviation is given, and α takes the same values ​​as above; Two-way retention condition: Only retain those that simultaneously meet the conditions. The optical flow point; Dynamic update mechanism: The threshold is recalculated for each frame based on the current frame's optical flow statistical distribution.

[0011] In a preferred embodiment, the state inference module includes: The mapping model unit is configured as a long short-term memory network, which includes an input gate, a forget gate, an output gate, and a memory unit. The input is a sequence of temporal optical flow feature matrices, and the output is the predicted motion state of the occluded target. The long short-term memory network is trained under unobstructed or weakly obstructed conditions. The training data includes a sequence of visible light images acquired simultaneously, a sequence of optical flow features at the edge of a vapor cloud, and the target motion trajectory output by the target detection unit. The unobstructed condition is defined as follows: the proportion of the vapor cloud coverage area is less than the third weak obstruction threshold, and the value of the third weak obstruction threshold ranges from 0% to 5%.

[0012] In a preferred embodiment, the state inference module further includes: The blindness determination unit is configured to determine that the sensor is in a blind state when the confidence level of the steam cloud coverage exceeds a first threshold and the proportion of the steam cloud coverage area exceeds a second threshold; wherein, the confidence level of the steam cloud coverage is calculated by the probability map output by the deep semantic segmentation network, and the proportion of the steam cloud coverage area is calculated by the binary mask map output by the steam cloud edge detection unit, and the proportion of the steam cloud coverage area is specifically the ratio of the number of pixels with a value of 1 in the binary mask map to the total number of pixels; The trajectory fusion unit is configured to use a Kalman filter to smooth the predicted trajectory output by the long short-term memory network and generate a complete trajectory during the occlusion period.

[0013] In a preferred embodiment, the blindness determination unit is configured to perform the following determination conditions: The confidence level for detecting the vapor cloud is greater than a first threshold, which is between 0.7 and 0.9. The proportion of the vapor cloud coverage area is greater than the second threshold, which is 40% to 60%; The confidence score for target detection within the occluded area is below a third threshold for three consecutive frames, where the third threshold is between 0.2 and 0.4. When all three conditions above are met simultaneously, the sensor is determined to be in a state of blindness.

[0014] In a preferred embodiment, the identity inheritance module is configured to perform the following matching conditions: Location distance condition: The Euclidean distance between the endpoint of the completed trajectory and the center of the restored target detection box is less than the fourth threshold, which is dynamically calculated based on the target size and is 0.3 to 0.8 times the width of the target detection box; Time difference condition: The time difference between the end of the occlusion and the appearance of the restored target is less than the fifth threshold, which is the time interval corresponding to 5 to 15 frames, calculated at 30 frames / second; Appearance similarity condition: The cosine similarity between the appearance features of the recovered target and the appearance features of the target stored before occlusion is greater than a sixth threshold, which is between 0.6 and 0.8.

[0015] In a preferred embodiment, a vapor cloud type discrimination module is further included, configured as follows: Synchronously collect process data from the distributed control system, including pressure values, temperature values, gas concentration values, and valve status; The steam cloud morphology features are time-stamped with the process data of the distributed control system to construct a multimodal feature vector. The steam cloud morphology features include the steam cloud area change rate, edge sharpness, and diffusion speed. A graph neural network is used to model the equipment relationships between chemical plants. A multimodal feature vector is input, and the graph neural network outputs the steam cloud type discrimination result, which includes normal process emissions and abnormal dangerous leaks. When the judgment result indicates an abnormal and dangerous leak, a high-risk warning signal is generated.

[0016] In a preferred embodiment, a decision output module is further included, configured as follows: The completed trajectory and identity inheritance results are pushed to the security control screen for visual display. When the status inference module determines that there is a safety risk to the obscured target, it automatically activates the audible and visual alarm and triggers the emergency response procedure.

[0017] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention abandons the traditional monitoring approach that relies on the characteristics of the target itself. By extracting motion feature information from the edge region of the steam cloud and combining it with a long short-term memory network to learn the nonlinear mapping relationship between the optical flow at the edge of the steam cloud and the motion of the internal target, it realizes the inference of the motion state from the visible edge to the invisible target. This successfully solves the problem of the target being unable to be identified and the tracking being interrupted under the obstruction of the steam cloud, and greatly improves the continuity of monitoring.

[0018] This invention divides a vapor cloud region into a core region and an edge transition region. The optical flow in the core region is labeled as a noise source and removed using bidirectional dynamic thresholding. The optical flow in the edge transition region is then processed with median filtering, effectively removing isolated noise points. Compared to existing single noise suppression schemes, this combined noise suppression strategy is more targeted, adapts to the dynamic characteristics of vapor clouds, significantly improves the accuracy of optical flow feature extraction, and provides reliable data support for inferring the target's motion state.

[0019] This invention designs a spatiotemporal consistency matching scheme based on three conditions: location distance, time difference, and appearance similarity. The Euclidean distance between the completed trajectory endpoint and the restored target, the time difference between occlusion and restoration, and the cosine similarity of appearance features are all given clear quantitative standards. This ensures that after the occlusion dissipates, the target identity before occlusion can be accurately inherited to the restored target, effectively avoiding the tracking breakage problem caused by target ID switching after occlusion.

[0020] This invention adds a steam cloud type discrimination module, which synchronously collects process data from the distributed control system, aligns the steam cloud morphology characteristics with the process data using timestamps, constructs a multimodal feature vector, and combines graph neural networks to model the correlation between chemical plants, accurately distinguishing between normal process emissions and abnormal dangerous leaks. This solves the problem that existing technologies cannot distinguish the nature of steam clouds, significantly reduces the false alarm rate of abnormal leaks, and generates high-risk early warning signals, realizing the linkage between safety monitoring and emergency response. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0022] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1, please refer to Figure 1 As shown in this embodiment, the AI-based intelligent security central control event analysis and decision support system includes: The occlusion recognition module is configured to receive the monitoring video stream and identify dynamic occlusion objects and their coverage areas in the video frame. The dynamic occlusion objects are non-rigid visual occlusion media that evolve dynamically over time. The motion feature extraction module is configured to extract motion feature information of the edge region of the dynamic occlusion when the target features in the video image cannot be normally identified due to the dynamic occlusion. The motion feature information refers to the optical flow vector field of the pixels on the edge contour of the steam cloud, where each optical flow vector contains pixel position coordinates, velocity vector, orientation angle, and amplitude information. The steam cloud edge region is the set of pixels determined by extracting its contour after identifying the steam cloud coverage area using a deep semantic segmentation network. The targets include personnel, equipment, vehicles, and open flames in a chemical scene. The state inference module is configured to infer the motion state of the occluded target during the occlusion period based on the motion feature information, and generate a complete trajectory during the occlusion period. The identity inheritance module is configured to match the completed trajectory with the restored target after the occlusion dissipates, thereby achieving target identity inheritance.

[0025] This invention abandons the traditional monitoring approach that relies on the target's own characteristics. By extracting motion feature information from the edge region of a steam cloud and combining it with a long short-term memory network to learn the nonlinear mapping relationship between the optical flow at the edge of the steam cloud and the motion of the target inside, it achieves motion state inference from the visible edge to the invisible target. This successfully solves the problem of target unidentification and tracking interruption under steam cloud occlusion, significantly improving the continuity of monitoring. By dividing the steam cloud region into a core region and an edge transition region, the optical flow in the core region is marked as a noise source and removed using bidirectional dynamic threshold segmentation, while the optical flow in the edge transition region is processed by median filtering, effectively removing isolated noise points. Compared with existing single noise suppression schemes, this combined noise suppression strategy is more targeted, adapts to the dynamic characteristics of steam clouds, significantly improves the accuracy of optical flow feature extraction, and provides reliable data support for target motion state inference.

[0026] This invention designs a three-condition spatiotemporal consistency matching scheme based on location distance, time difference, and appearance similarity. Clearly quantified standards are set for the Euclidean distance between the trajectory endpoint and the restored target, the time difference between occlusion and restoration, and the cosine similarity of appearance features. This ensures that after occlusion dissipates, the target identity before occlusion can be accurately inherited to the restored target, effectively avoiding tracking breakage caused by target ID switching after occlusion. A new steam cloud type discrimination module is added, synchronously collecting process data from the distributed control system. The steam cloud morphology features are timestamped with the process data to construct a multimodal feature vector. Combined with graph neural network modeling of chemical plant correlations, this accurately distinguishes between normal process emissions and abnormal dangerous leaks, solving the problem of existing technologies being unable to distinguish the nature of steam clouds. The false alarm rate for abnormal leaks is significantly reduced, and high-risk early warning signals are generated, achieving linkage between safety monitoring and emergency response.

[0027] In one embodiment, the dynamic occlusion is a steam cloud in a chemical industry setting, and the motion feature extraction module includes: The steam cloud edge detection unit is configured to use a deep semantic segmentation network to identify the steam cloud coverage area and extract edge contours. The deep semantic segmentation network adopts a U-Net architecture, introducing an attention mechanism at skip connections to enhance edge feature extraction capabilities. It is pre-trained on a chemical scene steam cloud image dataset (containing 5000 labeled images), with an input image size of 1920×1080 and an output binary mask of the same size, where 1 represents a steam cloud region and 0 represents a non-steam cloud region.

[0028] The network was pre-trained on a dataset of steam cloud images from chemical industry scenes (containing 5000 labeled images). The input image size is 1920×1080, and the output is a binary mask of the same size, where 1 represents a steam cloud region and 0 represents a non-steam cloud region. The Canny edge detection operator is used to extract the steam cloud edge contours from the binary mask image, with high and low thresholds set to 100 and 200, respectively, to obtain a set of edge pixels. .

[0029] The optical flow calculation unit is configured to use an improved Lucas-Kanade optical flow method based on an image pyramid structure to calculate the optical flow vector field of the pixels at the edge contour of the vapor cloud.

[0030] The specific implementation of the improved Lucas-Kanade optical flow method includes: constructing an image pyramid with L=4 layers, the original image being layer 0, and the image of layer l being obtained by downsampling the image of layer l-1 after Gaussian filtering (σ=0.5); calculating the optical flow starting from the top of the pyramid, using the calculation results of the upper layer as the initial estimate of the lower layer, and propagating downwards layer by layer; solving the optical flow equation for each layer using the standard Lucas-Kanade method, which is based on the assumptions of constant brightness, spatial consistency, and small motion. The optical flow equation is: ,in The gradient of the image in the x-direction, Let be the gradient of the image in the y-direction. The gradient of the image in the time direction, u is the component of the optical flow vector in the x direction (unit: pixels / frame), and v is the component of the optical flow vector in the y direction (unit: pixels / frame). The optical flow equation is solved by weighted least squares method, with a window size of 15×15. The weight of each pixel in the window is a Gaussian kernel function with a kernel standard deviation σ=2.0.

[0031] Optical flow feature construction unit, configured to construct optical flow feature matrix Each optical flow feature Includes pixel position coordinates (x, y) and velocity vector 、( This represents the component of velocity in the x-direction. (representing the component of velocity in the y-direction) and direction angle Amplitude Information. The optical flow feature matrix has a dimension of N×6, where N is the number of edge pixels. In this embodiment, N is a fixed value of 200 (if less, zeros are added; if more, random sampling is performed).

[0032] In one embodiment, the motion feature extraction module further includes an optical flow noise suppression unit, configured to: divide the vapor cloud region into a core region and an edge transition region, wherein the core region is a pixel value saturation region (grayscale value > 250), and the edge transition region is a gradient change band extending outward from the core region by 15 pixels; mark the optical flow in the core region as a noise source, and use bidirectional dynamic threshold segmentation to remove the optical flow noise in the core region; perform median filtering on the optical flow data retained in the edge transition region, with a filtering window size of 5×5, to remove isolated noise points.

[0033] The specific implementation of the bidirectional dynamic threshold segmentation is as follows: Horizontal segmentation: Threshold segmentation is performed on the component u of the optical flow in the x-direction, and the threshold value is... ,in This is a dynamic threshold in the horizontal direction, measured in pixels per frame. The mean of the u-component. The standard deviation is α, which is an adjustable sensitivity coefficient (ranging from 1.5 to 3.0; a larger value indicates stricter noise removal, while a smaller value indicates more edge motion information is retained). The value of α ranges from 1.5 to 3.0; in this embodiment, α = 2.0 is used. The determination of the above-mentioned range of α values ​​is based on statistical experiments using 500 frames of optical flow data from the edge of a steam cloud to determine the optimal range of α. When α < 1.5, the optical flow noise removal is insufficient, and the signal-to-noise ratio of the optical flow features is below 3dB; when α > 3.0, the effective edge motion information loss rate exceeds 15%; when α = 2.0, the signal-to-noise ratio reaches the optimal value of 8.5dB, and the effective information retention rate is greater than 90%. Compared with the existing single threshold segmentation, this bidirectional dynamic threshold segmentation method can more accurately distinguish between the noise optical flow in the core region of the steam cloud and the effective motion optical flow in the edge transition region, thereby significantly improving the reliability of optical flow feature extraction.

[0034] Vertical segmentation: Threshold segmentation is performed on the optical flow component v in the y-direction, with the threshold value... ,in This is a dynamic threshold in the vertical direction, measured in pixels per frame. The mean of the v component. The mean and standard deviation are the same as above; both the mean and standard deviation are calculated based on the optical flow v component of all vapor cloud edge contour pixels in the current frame. Two-way retention condition: Only retain those that simultaneously meet the conditions. Only select the optical flow points, and discard the rest; Dynamic update mechanism: The threshold is recalculated for each frame based on the current frame's optical flow statistical distribution.

[0035] The formulas for calculating the mean and standard deviation are as follows: ; ; Where N is the total number of pixels representing the edge contour of the steam cloud in the current frame. , These are the components of the optical flow vector of the i-th edge pixel in the x and y directions, respectively.

[0036] In one embodiment, the state inference module includes: The mapping model unit is configured as a Long Short-Term Memory (LSTM) network, which includes an input gate, a forget gate, an output gate, and a memory unit. The input is a sequence of temporal optical flow feature matrices. ,in, This represents the optical flow feature matrix at time t (the current frame). The time window length of the LSTM is the number of historical input frames. In this embodiment, it is taken as... This means that the target position in the current frame is predicted using the optical flow feature sequence of the past 10 frames.

[0037] The structural parameters of the LSTM network are as follows: the input dimension is N×6, where N=200 (number of optical flow feature points), and 6 is the feature dimension of each optical flow feature (position coordinates x, y, velocity components u, v, orientation angle θ, amplitude ||v||). This represents the direction angle of the optical flow vector, in radians, and its value range is... The magnitude of the optical flow vector, i.e., its velocity, is expressed in pixels per frame. The optical flow feature matrix has a dimension of N×6, where N=200, representing the number of feature points selected from the edge contour of the steam cloud. If the number is less than 200, it is padded with zeros; otherwise, random sampling is used. The LSTM has two layers, each with 128 hidden units. The output layer is a fully connected layer with a dimension of 2, corresponding to the coordinates of the predicted target center point. ,in The predicted x-axis, This represents the predicted ordinate.

[0038] The long short-term memory network is trained under unobstructed or weakly obstructed conditions. The training data includes synchronously acquired visible light image sequences, vapor cloud edge optical flow feature sequences, and target motion trajectories output by the target detection unit.

[0039] The training dataset contains 20,000 samples, each consisting of a 15-frame image sequence, divided into training, validation, and test sets in a 7:2:1 ratio. The Adam optimizer is used during training with a learning rate of 0.001, a batch size of 32, and a dropout rate of 0.2 to prevent overfitting. The mean squared error (MSE) loss function is used. ,in, The mean squared error loss value is used to measure the difference between the predicted trajectory and the actual trajectory. This represents the coordinates of the true center point of the target output by the target detection unit. , This represents the coordinates of the target center point predicted by the Long Short-Term Memory network. , The batch size is defined as follows. By minimizing this loss function, the network learns the mapping from the optical flow at the edge of the steam cloud to the motion state of the occluded target. The training epochs are 100, and an early stopping mechanism monitors the validation set loss; training stops if the loss does not decrease for 10 consecutive epochs.

[0040] In this embodiment, the unobstructed condition is defined as follows: the proportion of steam cloud coverage is less than 5% (i.e., the third weak occlusion threshold is 5%), at which point there is basically no steam cloud interference in the image, and the target detection unit can work normally. The weak occlusion condition is defined as follows: the proportion of steam cloud coverage is between 10% and 20% (i.e., the first weak occlusion threshold is 10% and 20%), and the target detection confidence is greater than 0.7 (i.e., the second weak occlusion threshold is 0.7), at which point there are a small amount of steam clouds in the image, but the target can still be partially identified.

[0041] The threshold range was determined based on the following: Statistical analysis of 5000 frames of steam cloud images of chemical scenes showed that when the steam cloud coverage area was less than 10%, the average confidence level of target detection was higher than 0.8; when the coverage area was between 10% and 20%, the average detection confidence level was approximately 0.6-0.7; and when the coverage area exceeded 20%, the target detection confidence level dropped sharply to below 0.4. Therefore, this application uses a coverage area ratio of 10%-20% as the quantitative definition standard for weak occlusion conditions.

[0042] The blindness determination unit is configured as follows: Obtain the probability map P and binary mask map M output by the steam cloud edge detection unit. The probability map P is output by the deep semantic segmentation network (U-Net), and each pixel value... This indicates the probability that the pixel belongs to a steam cloud; the binary mask image M is obtained by binarizing P with a threshold of 0.5, that is, M(i,j)=1 if P(i,j)≥0.5, otherwise 0.

[0043] Calculate the confidence level of vapor cloud cover and the proportion of steam cloud coverage area , Take the maximum probability of vapor cloud of all pixels in the whole image, where ΣM(i,j) represents the total number of pixels with a value of 1 in the binary mask image, and H and W are the height and width of the image (in pixels).

[0044] When the steam cloud coverage confidence level The proportion of steam cloud coverage area is greater than the first threshold (0.8 in this embodiment). When the confidence level of the target detection within the occluded area exceeds the second threshold (50% in this embodiment), the sensor is determined to be in a blind state. Furthermore, if the confidence level of the target detection within the occluded area is below the third threshold (0.3 in this embodiment) for three consecutive frames, the blind state is further confirmed to avoid misjudgment.

[0045] The first threshold (confidence level of vapor cloud coverage) ranges from 0.7 to 0.9, and is set to 0.8 in this embodiment. The higher the threshold, the stricter the requirement for certainty regarding the existence of vapor clouds, thus avoiding the erroneous triggering of a blind state due to false detection.

[0046] The second threshold (percentage of steam cloud coverage): the value ranges from 40% to 60%, and in this embodiment it is set to 50%. When the steam cloud covers more than half of the image, it is considered sufficient to cause severe visual obstruction, making it impossible to properly identify the target features.

[0047] The coverage area ratio is the core indicator for judging the blind state. Its physical meaning is that when the steam cloud covers more than half of the screen (i.e., the area ratio > 50%), the amount of effective information in the monitored field of view is lower than the threshold, and the confidence of the traditional target detection algorithm drops below 0.2. At this time, the blind state is judged and the LSTM inference logic is activated. This avoids the accuracy loss caused by activating the inference model too early, and ensures that the system can continue to output the target status when vision fails.

[0048] The trajectory fusion unit is configured to, during sensor blindness, fuse the optical flow feature matrix of the vapor cloud edge for each frame. The data is sequentially input into the Long Short-Term Memory (LSTM) network, and the network outputs the predicted coordinates of the center point of the occluded target frame by frame. The original predicted trajectory is formed; then, a Kalman filter is used to smooth the original predicted trajectory to generate the completed trajectory during the occlusion period. The state vector of the Kalman filter consists of the target position (x, y) and velocity. The observation vector is the predicted coordinates output by the Long Short-Term Memory network. The optimal estimated trajectory is obtained through a two-step recursive process of prediction and update.

[0049] The state equation of the Kalman filter is: The observation equation is: ,in, Let (x, y) be the target state vector at time k, where (x, y) are the coordinates of the target center point in pixels. Δt represents the velocity component, in pixels per second; A is the state transition matrix, A=[1,0,Δt,0;0,1,0,Δt;0,0,1,0;0,0,0,1], where Δt is the inter-frame time interval (1 / 30 second). Let Q be the process noise, which follows a Gaussian distribution N(0, Q) with mean 0, and Q be the process noise covariance matrix, denoted as . ; H is the observation vector, derived from the LSTM prediction result; H is the observation matrix; the observation matrix H=[1,0,0,0;0,1,0,0] is used to extract the target position coordinates as the observation vector; The observation noise follows a Gaussian distribution N(0, R) with a mean of 0, where R is the observation noise covariance matrix, denoted as . .

[0050] The prediction and update steps of Kalman filtering include: Prediction steps: ;in Let be the predicted state covariance matrix at time k (prior estimate covariance). Update steps: , ,in For Kalman gain, Let I be the estimated state covariance matrix at time k (posterior estimated covariance), and let I be the identity matrix with a dimension of 4×4.

[0051] In one embodiment, the identity inheritance module is configured to perform the following matching conditions: Location and distance conditions: Complete the trajectory endpoint Center of the recovered target detection box Euclidean distance between (Unit: pixels) is less than the fourth threshold, which is dynamically calculated based on the target size and is 0.5 times the width W of the target detection box. The preferred range of the fourth threshold is 0.3 to 0.8 times the width of the target detection box, which can be adjusted according to the accuracy requirements of the actual scene.

[0052] Time difference condition: the time when the occlusion ends When the recovery target appears The time difference between them is less than the fifth threshold, which is the time interval corresponding to 10 frames (i.e., 0.33 seconds, calculated at 30 frames / second); the preferred range of the fifth threshold is the time interval corresponding to 5 to 15 frames. Appearance similarity condition: The cosine similarity between the recovered target's appearance features and the target's stored appearance features before occlusion is greater than a sixth threshold, which is 0.7, preferably within the range of 0.6 to 0.8. The appearance features are fused using a color histogram (HSV space, 8×8×8 bins) and Local Binary Pattern (LBP) features. LBP uses a circular neighborhood sampling radius of 2 and 8 sampling points, resulting in a feature vector dimension of (512+256) = 768 dimensions. The cosine similarity calculation formula is: ,in Let A represent the cosine similarity, with a value range of [-1, 1]. Let A and B be the appearance feature vectors of the two targets, respectively.

[0053] When all three conditions are met, the identity of the target before occlusion is inherited to the restored target; if the match fails, a manual review mechanism is triggered, the restored target is treated as a new target and the abnormal event is recorded.

[0054] In one embodiment, a vapor cloud type discrimination module is also included, configured as follows: Synchronously collect process data from the distributed control system, including pressure values, temperature values, gas concentration values, and valve status. The sampling frequency is 10Hz, and the data is obtained in real time from the distributed control system via the OPCUA protocol, with a data delay of no more than 100ms. The steam cloud morphology characteristics are timestamped with the process data of the distributed control system to construct a multimodal feature vector. The steam cloud morphology characteristics include the steam cloud area change rate (…). The features reflect the diffusion rate, with ΔArea representing the change in vapor cloud area between adjacent frames and Area representing the vapor cloud area in the current frame. The ratio represents the relative rate of change. Other features include edge sharpness (mean gradient of edge pixels, reflecting droplet density) and diffusion rate (edge ​​expansion rate, per pixel / second). The feature dimensions are 3+4=7. The vapor cloud area change rate = ,in The area of ​​the vapor cloud in the current frame. This represents the area of ​​the steam cloud in the previous frame. A graph neural network (GNN) is used to model the equipment relationships between chemical plants. The GNN constructs a graph structure with plants as nodes and process pipelines as edges. Node features are distributed control system data (pressure, temperature, gas concentration, and valve status, totaling four dimensions). Edge weights are set according to the strength of process correlation, including material flow correlation (0-1), thermal coupling (0-1), and spatial proximity (0-1). A weighted average of these three factors yields the comprehensive weight. The probability of abnormal and dangerous leak categories output by the GNN ranges from [0,1], with values ​​closer to 1 indicating a higher confidence level in the model's judgment of an abnormal leak. The GNN employs a Graph Convolutional Network (GCN) architecture with 3 layers, 64 hidden units per layer, and ReLU activation. The input is a 7-dimensional multimodal feature vector, and the output is a 2-dimensional classification result (normal process emissions / abnormal hazardous leaks). The model training data comes from 5000 labeled steam cloud event samples, including 3000 normal emissions and 2000 abnormal leaks. The training set:validation set:test set ratio is 7:2:1. The cross-entropy loss function is used, with an Adam optimizer learning rate of 0.001, and the training run consists of 200 epochs.

[0055] When the probability of an abnormal dangerous leak category output by the graph neural network in the vapor cloud type discrimination module is greater than a preset threshold (0.7 in this embodiment), the discrimination result is an abnormal dangerous leak, and a high-risk warning signal is generated. After receiving the signal, the decision output module automatically activates the audible and visual alarm and triggers the emergency response process, including: automatically sending a work order to the handheld terminal of the nearest security personnel, the work order containing the target location, target identity, and risk type information; unlocking the access control of the surrounding escape routes; activating the emergency broadcast system and playing preset emergency voice messages; and simultaneously pushing information such as the leak location, leak level, and suggested evacuation routes to the security control screen, and pushing real-time images to the emergency command center screen.

[0056] In one embodiment, a decision output module is also included, configured as follows: The completed trajectory and identity inheritance results are pushed to the security central control screen for visualization. The completed trajectory is marked with a red dotted line on the digital twin map, and the actual recovered trajectory is marked with a green solid line. At the same time, information such as target identity ID, occlusion duration, and prediction confidence are also marked.

[0057] When the status inference module determines that the obscured target poses a safety risk (such as the target being within 5 meters of a leak source, the target being in a dangerous area, or the target's movement being abnormal, such as falling to the ground), it automatically activates the audible and visual alarm and triggers the emergency response process, including: automatically sending a work order to the nearest security personnel's handheld terminal (including the target's location, target's identity, and risk type); unlocking access control for surrounding escape routes; activating the emergency broadcast system (with preset emergency voice); and simultaneously pushing real-time video to the emergency command center's large screen.

[0058] This embodiment of the system underwent a 30-day field test in the ethylene plant area of ​​a chemical industrial park. The test scenarios included high-pressure steam discharge (safety valve depressurization) and cryogenic medium leakage simulation. The test results are as follows:

[0059] The test results verified that the system can effectively maintain target detection capability under extreme conditions where the signal-to-noise ratio approaches 0dB, significantly improving the reliability and intelligence level of the chemical safety control system.

[0060] 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.

Claims

1. An AI-based intelligent security central control event analysis and decision support system, characterized in that, include: The occlusion recognition module is configured to receive the monitoring video stream and identify dynamic occlusion objects and their coverage areas in the video frame. The dynamic occlusion objects are non-rigid visual occlusion media that evolve dynamically over time. The motion feature extraction module is configured to extract motion feature information of the edge region of the dynamic occlusion when the target features in the video frame cannot be properly identified due to the dynamic occlusion. The motion feature extraction module includes: The vapor cloud edge detection unit is configured to use a deep semantic segmentation network to identify the vapor cloud coverage area and extract the edge contour; The optical flow calculation unit is configured to use an improved Lucas-Kanade optical flow method based on an image pyramid structure to calculate the optical flow vector field of the pixels at the edge contour of the vapor cloud. The optical flow feature construction unit is configured to construct an optical flow feature matrix, where each optical flow feature contains pixel position coordinates, velocity vector, orientation angle, and amplitude information; The target detection unit is configured to perform target detection on a visible light image under unobstructed or weakly obstructed conditions and output the target motion trajectory; the motion feature information includes the optical flow vector field of the edge contour pixels of the dynamic occluder, and the target includes personnel, equipment, vehicles and open flames in a chemical scene; The state inference module is configured to infer the motion state of the occluded target during the occlusion period based on the motion feature information, and generate a completed trajectory during the occlusion period; the state inference module includes: The mapping model unit is configured as a long short-term memory network, with the input being a sequence of temporal optical flow feature matrices and the output being the predicted motion state of the occluded target; The blindness determination unit is configured to determine that the sensor is in a blind state when the confidence level of the vapor cloud coverage exceeds a first threshold and the proportion of the vapor cloud coverage area exceeds a second threshold. The trajectory fusion unit is configured to use a Kalman filter to smooth the predicted trajectory output by the long short-term memory network and generate a complete trajectory during the occlusion period. The identity inheritance module is configured to match the completed trajectory with the restored target after the occlusion dissipates, thereby achieving target identity inheritance. The identity inheritance module is configured to perform the following matching conditions: Location distance condition: The Euclidean distance between the endpoint of the completed trajectory and the center of the restored target detection box is less than the fourth threshold; Time difference condition: The time difference between the end of occlusion and the appearance of the restored target is less than the fifth threshold; Appearance similarity condition: The cosine similarity between the appearance features of the recovered target and the appearance features of the target stored before occlusion is greater than the sixth threshold.

2. The AI-based intelligent security central control event analysis and decision support system according to claim 1, characterized in that: The dynamic occlusion is specifically a steam cloud in a chemical industry setting, and the motion feature extraction module includes: The vapor cloud edge detection unit is configured to use a deep semantic segmentation network to identify the vapor cloud coverage area and extract the edge contour; The optical flow calculation unit is configured to use an improved Lucas-Kanade optical flow method based on an image pyramid structure to calculate the optical flow vector field of the pixels at the edge contour of the vapor cloud. The optical flow feature construction unit is configured to construct an optical flow feature matrix, where each optical flow feature contains pixel position coordinates, velocity vector, orientation angle, and amplitude information; The target detection unit is configured to perform target detection on a visible light image under unobstructed or weakly obstructed conditions and output the target motion trajectory. The weak occlusion condition is defined as follows: the proportion of the vapor cloud coverage area is less than the first weak occlusion threshold, and the target detection confidence is greater than the second weak occlusion threshold; the value range of the first weak occlusion threshold is 10% to 20%, and the value range of the second weak occlusion threshold is 0.6 to 0.

8.

3. The AI-based intelligent security central control event analysis and decision support system according to claim 1, characterized in that: The motion feature extraction module further includes an optical flow noise suppression unit, configured as follows: The vapor cloud region is divided into a core region and an edge transition region. The core region is the pixel value saturation area, and the edge transition region is the gradient change zone extending outward from the core region. The optical flow in the core region is marked as a noise source, and bidirectional dynamic threshold segmentation is used to remove optical flow noise in the core region. Median filtering is applied to the optical flow data retained in the edge transition region to remove isolated noise points.

4. The AI-based intelligent security central control event analysis and decision support system according to claim 3, characterized in that: When the optical flow noise suppression unit is configured to perform the bidirectional dynamic thresholding, it specifically performs the following: Horizontal segmentation: Threshold segmentation is performed on the optical flow component u in the x-direction, with the threshold value... ,in The mean of the u-component. The standard deviation is denoted by α, which is an adjustable coefficient ranging from 1.5 to 3.

0. Vertical segmentation: Threshold segmentation is performed on the optical flow component v in the y-direction, with the threshold value... ,in The mean of the v component. The standard deviation is given, and α takes the same values ​​as above; Two-way retention condition: Only retain those that simultaneously meet the conditions. The optical flow point; Dynamic update mechanism: The threshold is recalculated for each frame based on the current frame's optical flow statistical distribution.

5. The AI-based intelligent security central control event analysis and decision support system according to claim 1, characterized in that: The Long Short-Term Memory (LSTM) network includes an input gate, a forget gate, an output gate, and memory units; The long short-term memory network is trained under unobstructed or weakly obstructed conditions. The training data includes a sequence of visible light images acquired simultaneously, a sequence of optical flow features at the edge of a vapor cloud, and the target motion trajectory output by the target detection unit. The unobstructed condition is defined as follows: the proportion of the vapor cloud coverage area is less than the third weak obstruction threshold, and the value of the third weak obstruction threshold ranges from 0% to 5%.

6. The AI-based intelligent security central control event analysis and decision support system according to claim 5, characterized in that: The confidence score of the steam cloud coverage is calculated from the probability map output by the deep semantic segmentation network, and the proportion of the steam cloud coverage area is calculated from the binary mask map output by the steam cloud edge detection unit. Specifically, the proportion of the steam cloud coverage area is the ratio of the number of pixels with a value of 1 in the binary mask map to the total number of pixels.

7. The AI-based intelligent security central control event analysis and decision support system according to claim 1, characterized in that: The blindness determination unit is configured to perform the following determination conditions: The confidence level for detecting the vapor cloud is greater than a first threshold, which is between 0.7 and 0.

9. The proportion of the vapor cloud coverage area is greater than the second threshold, which is 40% to 60%; The confidence score for target detection within the occluded area is below a third threshold for three consecutive frames, where the third threshold is between 0.2 and 0.

4. When all three conditions above are met simultaneously, the sensor is determined to be in a state of blindness.

8. The AI-based intelligent security central control event analysis and decision support system according to claim 1, characterized in that: The fourth threshold is dynamically calculated based on the target size and is 0.3 to 0.8 times the width of the target detection box. The fifth threshold is the time interval corresponding to 5 to 15 frames, calculated at 30 frames / second. The sixth threshold is between 0.6 and 0.

8.

9. The AI-based intelligent security central control event analysis and decision support system according to claim 1, characterized in that: It also includes a steam cloud type discrimination module, configured as follows: Synchronously collect process data from the distributed control system, including pressure values, temperature values, gas concentration values, and valve status; The steam cloud morphology features are time-stamped with the process data of the distributed control system to construct a multimodal feature vector. The steam cloud morphology features include the steam cloud area change rate, edge sharpness, and diffusion speed. A graph neural network is used to model the equipment relationships between chemical plants. A multimodal feature vector is input, and the graph neural network outputs the steam cloud type discrimination result, which includes normal process emissions and abnormal dangerous leaks. When the judgment result indicates an abnormal and dangerous leak, a high-risk warning signal is generated.

10. The AI-based intelligent security central control event analysis and decision support system according to claim 1, characterized in that: It also includes a decision output module, configured as follows: The completed trajectory and identity inheritance results are pushed to the security control screen for visual display. When the status inference module determines that there is a safety risk to the obscured target, it automatically activates the audible and visual alarm and triggers the emergency response procedure.

Images