Hidden danger model generation method and device based on intelligent identification and spatial grid algorithm

Abstract

This invention discloses a method and apparatus for generating hazard models based on intelligent recognition and spatial grid algorithms. The method includes: preprocessing and differentially labeling an initial image set to obtain a structured labeled dataset, then training a corresponding hazard recognition neural network; identifying historical hazard image sets using the hazard recognition neural network to obtain hazard identification results and target risk intervals for hazard areas; constructing a corresponding hazard scene model and inferring a spatial Gaussian scene; dynamically adjusting the basic risk weights of key areas in the hazard scene model and matching the spatial Gaussian scene with the hazard scene model to obtain a corresponding target hazard scene model. This method automatically generates simulation scene models corresponding to real-world hazards through virtual simulation and 3D interactive technology. Users can intuitively learn hazard characteristics and practice / assess in a virtual environment, improving the efficiency of user training and enhancing the immersion and practicality of the training.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for generating hazard models based on intelligent recognition and spatial grid algorithms. Background Technology

[0002] Traditional hazard identification methods rely primarily on manual inspections and experience-based judgment, resulting in low identification efficiency, highly subjective results, and difficulty in reproducing hazards. Meanwhile, safety training for hazard identification is often disconnected from hazard identification, remaining at the stage of static video playback and centralized lectures, lacking connection to real-world scenarios. This leads to limited training effectiveness, low learning enthusiasm, and poor learning outcomes. However, existing technical methods for safety training in hazard identification also suffer from these problems. Summary of the Invention

[0003] This invention provides a method and apparatus for generating hazard models based on intelligent identification and spatial grid algorithms, aiming to solve the problem of poor training effectiveness in existing safety training methods for hazard identification. By applying the target hazard scenario model generated by the hazard model generation method in this solution for training, the efficiency and practicality of safety training for hazard identification can be greatly improved.

[0004] In a first aspect, embodiments of the present invention provide a method for generating a hazard model based on intelligent identification and spatial grid algorithms, the method comprising: Receive the input initial image set, and preprocess the images contained in the initial image set to obtain the corresponding preprocessed image set; The images in the preprocessed image set are differentially labeled according to the preset differential labeling rules to obtain the corresponding structured labeled dataset; The pre-stored initial identification network is trained based on the structured labeled dataset to obtain the corresponding hazard identification neural network; The system receives a set of historical hazard images and identifies them using the hazard identification neural network to obtain the corresponding hazard identification results. Based on the preset risk probability analysis rules and the hazard identification results, risk analysis is performed on each hazard area in the historical hazard image set to obtain the target risk interval corresponding to each hazard area. Based on the feature information corresponding to the historical hazard image set, the target risk range, and the risk probability analysis information, a corresponding hazard scenario model is constructed. Reasoning is performed based on the aforementioned potential hazard scenario model to generate a corresponding spatial Gaussian scenario; Based on the spatial scene characteristics of the Gaussian spatial scene, the basic risk weights of key areas in the hidden danger scene model are dynamically adjusted to generate a corresponding comprehensive risk weight set. The spatial Gaussian scene and the hidden danger scene model are matched according to the preset spatial grid matching algorithm and the comprehensive risk weight set to obtain the corresponding target hidden danger scene model.

[0005] Secondly, embodiments of the present invention also provide a hazard model generation apparatus based on intelligent identification and spatial grid algorithms, wherein the apparatus is used to execute the hazard model generation method based on intelligent identification and spatial grid algorithms as described in the first aspect above, and the apparatus includes: The image preprocessing unit is used to receive the input initial image set and preprocess the images contained in the initial image set to obtain the corresponding preprocessed image set. The annotation unit is used to perform differential annotation on the images contained in the preprocessed image set according to the preset differential annotation rules to obtain the corresponding structured annotation dataset; The network training unit is used to train the pre-stored initial identification network based on the structured labeled dataset to obtain the corresponding hidden danger identification neural network. The identification unit is used to receive the input set of historical hazard images, and to identify the set of historical hazard images according to the hazard identification neural network to obtain the corresponding hazard identification result; The risk analysis unit is used to perform risk analysis on each hidden danger area in the historical hidden danger image set according to the preset risk probability analysis rules and the hidden danger identification results, and to obtain the target risk interval corresponding to each hidden danger area. The scenario model construction unit is used to construct a corresponding hazard scenario model based on the feature information corresponding to the historical hazard image set, the target risk range, and the risk probability analysis information. The scene generation unit is used to perform reasoning based on the hidden danger scene model to generate a corresponding spatial Gaussian scene; The weight adjustment unit is used to dynamically adjust the basic risk weights of key areas in the hidden danger scenario model according to the spatial scene characteristics of the spatial Gaussian scene, so as to generate a corresponding comprehensive risk weight set. The target hazard scene model acquisition unit is used to match the spatial Gaussian scene with the hazard scene model according to the preset spatial grid matching algorithm and the comprehensive risk weight set, so as to obtain the corresponding target hazard scene model.

[0006] Thirdly, embodiments of the present invention also provide a computer device, wherein the device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of the hazard model generation method based on intelligent identification and spatial grid algorithm described in the first aspect above.

[0007] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the hazard model generation method based on intelligent identification and spatial grid algorithm described in the first aspect above.

[0008] This invention provides a method and apparatus for generating hazard models based on intelligent recognition and spatial grid algorithms. The method includes: preprocessing an initial image set and differentially labeling it to obtain a structured labeled dataset, then training a corresponding hazard recognition neural network; using the hazard recognition neural network to identify historical hazard image sets to obtain hazard identification results and acquire the target risk interval of the hazard area; constructing a corresponding hazard scene model and inferring a spatial Gaussian scene; dynamically adjusting the basic risk weights of key areas in the hazard scene model and matching the spatial Gaussian scene with the hazard scene model to obtain the corresponding target hazard scene model. This hazard model generation method automatically generates a simulation scene model corresponding to real-world hazards through virtual simulation and 3D interactive technology. Users can intuitively learn hazard characteristics and practice / assess in a virtual environment, improving the efficiency of user training and enhancing the immersion and practicality of the training. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating the method for generating a hazard model based on intelligent identification and spatial grid algorithms provided in an embodiment of the present invention. Figure 2 A schematic block diagram of a hazard model generation device based on intelligent identification and spatial grid algorithm provided in an embodiment of the present invention; Figure 3 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0011] 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, not all, of the embodiments of the present invention. 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.

[0012] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0013] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0014] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0015] The embodiments of this invention application provide a method for generating a hazard model based on intelligent identification and spatial grid algorithms. This method is applied to a terminal device, and the hazard model generation method is executed by application software installed on the terminal device. The terminal device is the terminal device that executes the hazard model generation method to generate a corresponding target hazard scene model for training and learning, such as a desktop computer, laptop computer, tablet computer, or mobile phone.

[0016] like Figure 1 As shown, the method includes steps S110 to S190.

[0017] S110. Receive the input initial image set, and preprocess the images contained in the initial image set to obtain the corresponding preprocessed image set.

[0018] Users take multi-angle photos of the factory area covering various potential hazard scenarios to form an initial image set. This initial image set is then input into a terminal device, which preprocesses the images within the set to obtain a preprocessed image set. The initial image set synchronously records the factory area type tag corresponding to each image.

[0019] Specifically, image preprocessing involves automatically filtering out low-quality images such as blurry or overexposed images using algorithms, and then performing illumination normalization and lens distortion correction on the remaining images. The automatic filtering rules include first checking the image resolution, determining if the minimum width and minimum height meet the minimum resolution requirements (checking if the minimum width is greater than a width threshold and if the minimum height is greater than a height threshold); images that do not meet the minimum resolution requirements are filtered out. Further, blur detection is performed on the remaining images, using the Laplacian operator to calculate image sharpness. A higher variance value for the Laplacian operator indicates a sharper image. The image's Laplacian operator variance value is checked against a variance threshold; if it is not, the image is filtered out.

[0020] Calculate the average brightness of the image and the standard deviation of the image's contrast. If the average brightness of the image is lower than the brightness threshold, or the standard deviation of the image's contrast is lower than the standard deviation threshold, the image is filtered out to exclude images that are too dark or have too low contrast.

[0021] Furthermore, overexposure detection can be performed on the remaining images. The proportion of high-brightness pixels (pixel grayscale values ​​greater than 245 are considered high-brightness pixels) in the image can be used as the overexposure pixel ratio. If the overexposure pixel ratio is greater than a threshold, an overexposure region mask is generated. Based on the overexposure region mask, morphological operations are performed on the overexposure regions to enhance connectivity. ConnectedComponentsWithStats is used to analyze connected components and calculate the proportion of the largest white region to the image area. If the proportion of the connected component to the image area is greater than a preset proportion value, the image is filtered out.

[0022] S120. Perform differential annotation on the images contained in the preprocessed image set according to the preset differential annotation rules to obtain the corresponding structured annotation dataset.

[0023] Furthermore, differential annotation is performed on the images in the obtained preprocessed image set. Differential annotation can be achieved through differential annotation rules. By annotating the images in the preprocessed image set, a structured labeled dataset containing bounding boxes, polygons, and key point coordinates can be obtained.

[0024] In a specific embodiment, step S120 includes the following sub-steps: classifying the images in the preprocessed image set according to the classification types included in the differential annotation rules to obtain the classification image sets corresponding to each classification type; annotating the classification image sets corresponding to the point hazard type according to the point hazard annotation format in the differential annotation rules to obtain the corresponding point hazard annotation information; annotating the classification image sets corresponding to the regular hazard type according to the regular hazard annotation format in the differential annotation rules to obtain the corresponding regular hazard annotation information; annotating the classification image sets corresponding to the irregular hazard type according to the irregular hazard annotation format in the differential annotation rules to obtain the corresponding irregular hazard annotation information; and integrating the point hazard annotation information, the regular hazard annotation information, and the irregular hazard annotation information to obtain the corresponding structured annotation dataset.

[0025] The images are classified according to the morphological type of the hazards contained in the preprocessed image set. The morphological type of the hazard is also a classification type in the differential labeling rules. Differential labeling schemes are used to label the hazards according to their visual morphology. The classification types include point-shaped hazard type, regular hazard type, and irregular hazard type.

[0026] The images in the preprocessed image set can be classified according to the classification types included in the differentiated annotation rules, thus obtaining the corresponding classification image sets for each classification type. Further, the images in the classification image sets corresponding to point-like hazard types (such as bolt loosening, weld defects) are annotated according to the point-like hazard annotation format. This means using a single key point for annotation, with the annotation content being the (x, y) coordinates of the hazard's center point. These coordinates are the key point coordinates for annotation; one or more key point coordinates can be annotated for a single image. This step then yields the point-like hazard annotation information corresponding to the classification image sets.

[0027] According to the rule-based hazard labeling format, images in the corresponding classification image set for rule-based hazard types (such as cracks and regular corrosion zones) are labeled using rectangular boxes. The labeling content is the coordinates (x1, y1, x2, y2) of the top-left and bottom-right corners of the rectangle. One image can be labeled with one or more rectangles. This step then yields the rule-based hazard labeling information corresponding to the classification image set.

[0028] According to the irregular hazard labeling format, images in the corresponding classification image set for irregular hazard types (such as paint peeling, irregular damage) are labeled. This is done using polygon labeling, where the labeling content is a list of coordinates [(x1,y1),(x2,y2),...] of the polygon vertices recorded in sequence. One image can be labeled with one or more polygon labels. This step then yields the irregular hazard labeling information corresponding to the classification image set.

[0029] The point-like hazard annotation information, regular hazard annotation information, and irregular hazard annotation information contained in the images of each image set are integrated to obtain the corresponding structured annotation dataset.

[0030] S130. Train the pre-stored initial identification network according to the structured labeled dataset to obtain the corresponding hidden danger identification neural network.

[0031] The initial identification network is trained using the obtained structured labeled dataset, resulting in a trained neural network used for hazard identification. Specifically, the initial identification network can be built based on a redesigned YOLOv8 architecture. The core of this initial identification network incorporates a multi-scale adaptive attention mechanism and a factory area type conditional fusion module (PF-Fusion). The convolutional processing of the multi-scale adaptive attention mechanism is as follows: Where S={1,3,5} represents different convolutional kernel sizes, X is the input image vector (feature vectors of images in a structured labeled dataset), and GAP is global average pooling. This represents element-wise multiplication. Features are enhanced by introducing a multi-scale adaptive attention mechanism, and the detection is adapted to specific scenarios by introducing a factory type conditional fusion module.

[0032] Multi-scale adaptive attention is woven into the critical path of feature extraction and fusion. When feature maps undergo downsampling in the backbone, this mechanism works like an embedded modulator: on shallow, high-resolution features, it quietly enhances the response to local details such as texture edges, guiding the network to notice small target traces that are easily overlooked; as the network deepens and feature maps shrink, it turns to capturing more abstract semantic context, automatically balancing the importance between different channels to ensure that key category information is not diluted.

[0033] This mechanism acts as an intelligent arbitrator during bidirectional fusion of feature pyramids. When deep semantic features propagate upwards and encounter shallow detail features, they are not simply added together. Instead, a set of weights is dynamically generated to determine whether more detailed location information or stronger semantic concepts should be trusted at a given spatial location. This adaptive fusion method ensures that the output at each scale contains just the right balance of detail and semantics, enabling subsequent detection heads to obtain the feature base with optimal representational power, regardless of whether they are processing large or small targets.

[0034] Throughout the process, attention weights are generated entirely by data, without the need for pre-defined fixed patterns. During training, the initial recognition network learns to automatically adjust its attention strategy based on different input scenarios—expanding its spatial focus when faced with densely packed small objects, and focusing on discriminative local features when encountering large targets. This dynamic decision-making process, inherent in the network's forward propagation, is the core mechanism behind its enhanced multi-scale generalization ability.

[0035] Using images from a structured, labeled dataset, the composite loss function L is minimized. λ 1L box + λ 2L poly + λ 3L point + λ 4L cls The parameters of the initial recognition network are trained end-to-end. Based on an active learning strategy, the training set is iteratively expanded using low-confidence detection results, and a lightweight version is developed to optimize the inference efficiency of the initial recognition network on different hardware.

[0036] The steps of expanding the training set using the pseudo-label method include: (1) Initial training of the network, initial recognition network M0 → in the structured labeled dataset D labeled Training → Obtain the neural network M1 after the first training; (2) Pseudo-label generation, ŷ=M1(X)={bbox,polygon,keypoints,class,confidence}, X is the feature vector of the image, bbox is the regular hazard labeling information corresponding to the image, polygon is the irregular hazard labeling information corresponding to the image, keypoints is the point hazard labeling information corresponding to the image, class is the classification type of the image, confidence is the confidence of the pseudo-label, and the selection condition is: τ low <confidence<τ high , where: τ low =0.3, τ high =0.7; (3) Multi-level filtering mechanism: First, the generated pseudo-labels are sorted based on confidence to obtain pseudo-label sorting information, and then non-maximum suppression (NMS) is performed. For any pseudo-label in the pseudo-label sorting information, if there is another pseudo-label with which the intersection-union ratio (IoU) is greater than the IoU threshold (e.g., set to 0.5), then the other pseudo-label is removed to obtain the initial pseudo-label set; then uncertainty screening is performed, and the classification entropy of the same classification type in the initial pseudo-label set is calculated, U(p)=H(p) class )=-∑p i ·log(p i ), where p iFor the i-th pseudo-label of a certain category in the candidate pseudo-label set, if the classification entropy U(p) of a certain category is less than the upper quartile (Q3(U)) of the classification entropy U of all samples, then the image corresponding to that category is retained, and the retained pseudo-label is used as the candidate pseudo-label set. Further, valid conditions are set to perform a geometric rationality check on the pseudo-labels included in the candidate pseudo-label set, such as setting the valid condition = {polygons do not self-intersect: is simple (polygon)=True, key points are inside the bounding box: And if the shape regularity score is greater than the threshold, obtain the pseudo-label combinations that meet the set conditions as the target pseudo-label set P. filtered (4) Training set expansion, D labeled =D labeled ∪P filtered (5) Iterative optimization: Repeat steps (1) to (4). In each iteration, train a new neural network M. k New pseudo-labels are generated to expand the dataset and update the neural network.

[0037] The iteration termination condition can be set to reach the maximum number of iterations K (usually K=5), or to stop improving the quality of pseudo-labels, or to convergence of the validation set performance. After training, the resulting neural network is the hazard identification neural network. After inference, the hazard identification neural network outputs standardized results after processing such as nonmaximum suppression. Example format: {"Hazard type":"Fire escape route blocked","Annotation shape":"Rectangle","Coordinates":[[x1,y1],[x2,y2]],"Confidence":0.95}, providing accurate 2D hazard information for the subsequent generation of 3D hazard scene models.

[0038] S140. Receive the input set of historical hazard images, and identify the set of historical hazard images according to the hazard identification neural network to obtain the corresponding hazard identification result.

[0039] The terminal device further receives the input historical hazard image set. This historical hazard image set can be obtained by adding historical hazard stimuli, safety accident reports, and equipment inspection logs associated with the area in the initial image set. Alternatively, a separate input historical hazard image set containing multi-view RGB images of the plant area, historical hazard stimuli, safety accident reports, and equipment inspection logs can be used. This historical hazard image set can also serve as a multimodal dataset of image-image location information-risk text information. The feature information corresponding to the image includes image location information and risk text information. The risk text information is presented as text-based feature description information, and the image location information records the location of the image.

[0040] Specifically, the original images in the historical hazard image set can be automatically screened for quality and subjected to geometric / photometric correction, and the associated risk text information can be structured to unify the standards and levels of risk description.

[0041] The geometric correction of the original image involves using a calibration board (such as a checkerboard or Charuco board) or self-calibration based on Structure of Motion (SFM). Specific steps include data acquisition, calculation of intrinsic parameters and distortion coefficients, image distortion correction, and intrinsic parameter consistency processing. The photometric correction of the original image involves global correction of the initial image during preprocessing to eliminate photometric variations not caused by the scene itself. Processing steps include exposure and white balance unification, high dynamic range processing, and lens vignetting correction.

[0042] Specifically, after quality screening and correction of the initial images in the historical hazard image set, feature vectors corresponding to each image are extracted from the processed images. The feature vectors of the images are then input into the aforementioned hazard identification neural network to obtain the corresponding hazard identification information. The obtained hazard identification information is as shown in the inference output results in the above steps. Thus, each image can obtain a corresponding hazard identification information, and the obtained hazard identification results contain the hazard identification information of each image.

[0043] S150. Based on the preset risk probability analysis rules and the hazard identification results, perform risk analysis on each hazard area in the historical hazard image set to obtain the target risk interval corresponding to each hazard area.

[0044] Furthermore, based on the risk probability analysis rules and the hazard identification results obtained from the above steps, risk analysis is performed on each hazard area in the historical hazard image set to obtain the target risk range corresponding to each hazard area.

[0045] In a specific embodiment, step S150 includes the following sub-steps: obtaining the experience interval of each hidden danger area by matching the risk weight range in the risk probability analysis rule according to each hidden danger area in the historical hidden danger image set; obtaining the hidden danger information corresponding to each hidden danger area in the hidden danger identification result; calculating the corresponding risk statistical index for the hidden danger information of each hidden danger area according to the risk value calculation formula in the risk probability analysis rule; mapping the risk statistical index to the experience interval of the corresponding hidden danger area to obtain the interval mapping value corresponding to each hidden danger area as the risk probability analysis information; and correcting the experience interval according to the observation data corresponding to the historical hidden danger image set and the risk probability analysis information to obtain the updated target risk interval for each hidden danger area.

[0046] Specifically, the risk probability analysis rules include multiple risk weight ranges. Based on the historical hazard image set after screening and correction, the risk weight range of each hazard area is matched from the risk probability analysis rules and used as the empirical interval corresponding to the hazard area.

[0047] For example, the risk weight ranges defined based on historical data are as follows: Power distribution room / hazardous chemical storage area: [0.75, 0.95] - high inherent risk, serious consequences of accidents; Equipment area (heavy machinery / high temperature and high pressure): [0.60, 0.85] - medium to high risk, strongly correlated with operation; Storage area (high-bay warehouse / turnover area): [0.40, 0.70] - medium risk, common hazards of falls and collisions; Production passage / main road: [0.30, 0.60] - medium to low risk, intersection of people and logistics; Office area / rest area: [0.10, 0.30] - low risk, relatively stable environment.

[0048] To further obtain the hazard information corresponding to each hazard area in the hazard identification results, it is possible to determine whether the images are in the same hazard area based on their location. If they are in the same hazard area, the hazard identification information of multiple images located in the same hazard area is combined to obtain the hazard information corresponding to that hazard area.

[0049] The risk probability analysis rules include a risk value calculation formula. This formula can be used to calculate the hazard information for each potential hazard area, thereby obtaining the corresponding risk statistical index. Specifically, the risk value calculation formula RI=α can be used... R ×F+β R The risk statistical index RI is calculated using ×T, with RI ranging from [0,1]; F represents the frequency of historical hazards / events in the hazard area (times / unit time), normalized to [0,1]; T represents the average severity of historical events in the hazard area (e.g., classified by lost working hours or economic losses), normalized to [0,1]; α R and β R Harmonic weights for frequency and severity (e.g., α can be set) R =0.4, β R =0.6). Then the risk statistical index corresponding to each risk interval can be calculated separately.

[0050] Furthermore, the risk statistical index is mapped to the empirical interval of the corresponding hidden danger area. The specific mapping process can be expressed as r base =R min +RI×(R max -R min ), R min R is the lower limit of the empirical interval corresponding to the potential hazard area. max r is the upper limit of the experience interval corresponding to the potential hazard area.base Let r be the interval mapping value obtained by mapping. base It retains the upper and lower bounds of expert experience while being driven by actual statistical data, performing linear interpolation within the interval to locate potential hazards. The interval mapping values ​​corresponding to each hazard area are obtained as risk probability analysis information.

[0051] Furthermore, based on the observation data corresponding to the historical hazard image set and the obtained risk probability analysis information, the empirical intervals of each hazard area are corrected to obtain the updated target risk intervals for each hazard area.

[0052] The empirical intervals are corrected based on the observation data corresponding to the historical hazard image set and the risk probability analysis information to obtain the updated target risk intervals for each hazard area. Specifically, this includes: obtaining observation feature information corresponding to the observation data; performing Bayesian updates on the empirical intervals of each hazard area based on the statistical features of the empirical intervals and the corresponding observation feature information to obtain the posterior distribution information corresponding to each hazard area; and obtaining the corresponding target intervals from the posterior distribution information of each hazard area based on a preset confidence level.

[0053] Observational data refers to the data feedback from users observing potential hazard areas. Specifically, it involves obtaining observational feature information corresponding to the observational data, and updating the prior distribution (the currently obtained empirical interval) to the posterior distribution (the corrected target interval) based on Bayesian updates. This observational feature information includes the suggested risk observation value r. obs (r) obs =r base +Δr,r base Here, Δr represents the interval mapping value for a specific hidden danger area, Δr is the user-suggested adjustment value, the feedback quality factor q (q≈0.9 for high-quality feedback (experts, detailed records), and q≈0.5 for general feedback), and the observation standard deviation σ. obs =(1-q)×σ prior .

[0054] The experience interval [R] for each potential hazard area min ,R max Assuming the risk value follows a truncated normal distribution, the statistical characteristics of each empirical interval can be obtained, with the mean μ. prior =(R min +R max ) / 2, standard deviation σ prior =(R max -R min ) / 4 (95% confidence level), cutoff interval [R] min ,R max ].

[0055] Each potential hazard area corresponds to a set of statistical features and a set of observational features. Based on these statistical and observational features, a Bayesian update is performed on the empirical interval of the potential hazard area to obtain its posterior distribution information. Specifically, N observations {r} of the potential hazard area within a time window are collected. obs-i ,q i}, calculate the corresponding valid observations: r obseff =Σ(w i ×r obs-i ) / Σw i Weight w i =1 / σ obs-i ²; Calculate the population observation variance: σ obs-total =1 / √(Σ(1 / σ obs-i ²)).

[0056] The Bayesian update formula is: posterior mean μ post =(μ prior / σ prior ²+r obs-eff / σ obs-total ²) / (1 / σ prior ²+1 / σ obs-total ²); Posterior variance σ post ²=1 / (1 / σ prior ²+1 / σ obs-total ²), then the obtained posterior mean and posterior variance are used as posterior distribution information. Based on the obtained posterior mean and posterior variance, a 95% confidence interval is obtained, thus yielding the corresponding target interval. The lower limit of the target interval is R. min-new =max(0,μ post -2σ post The upper limit of the target interval is R. max-new =min(1,μ post +2σ post ).

[0057] S160. Construct a corresponding hidden danger scenario model based on the feature information corresponding to the historical hidden danger image set, the target risk range, and the risk probability analysis information.

[0058] Furthermore, based on the feature information, target risk range, and risk probability analysis information corresponding to the historical hazard image set, a corresponding hazard scene model is constructed. A multi-view stereo algorithm is used to restore the accurate 3D scene structure. Based on the constructed 3D scene structure, combined with image semantic segmentation and point cloud clustering, the scene is automatically segmented into a discrete set of semantically meaningful "key regions" {R}. k}, for each region R k Associate it with its pre-calculated basic risk weight rbase-k .

[0059] In a specific embodiment, step S160 includes the following sub-steps: performing virtual modeling based on the image location information corresponding to each image in the historical hazard image set from the feature information to obtain a corresponding basic scene model; extracting corresponding point cloud feature vectors from each image in the historical hazard image set; performing clustering segmentation on the basic scene model according to preset hierarchical clustering segmentation rules and the point cloud feature vectors to obtain corresponding clusters; merging and optimizing the regions contained in the clusters according to preset merging rules to obtain a corresponding set of key regions; obtaining weighted risk coefficients corresponding to each segmented region in the set of key regions based on the point cloud feature vectors and feature description information in the feature information; normalizing the weighted risk coefficients of each segmented region according to the target risk interval to obtain the basic risk weights corresponding to each segmented region; and configuring the weights of the basic scene model according to the basic risk weights of each segmented region to obtain a corresponding hazard scene model.

[0060] Specifically, virtual modeling is performed based on the image location information corresponding to each image in the historical hazard image set after screening and correction, thereby constructing a 3D scene structure corresponding to the historical hazard image set as the basic scene model.

[0061] Further, point cloud feature vectors are extracted from each image in the historical hazard image set after screening and correction. Obtaining the point cloud feature vectors involves feature vector extraction and feature standardization. For each 3D Gaussian point in the image, the following feature vector F is extracted. i =[position(3),color(3),scale(3),rotation(4),local density (1)]; where position is the spatial coordinate (x, y, z), color is the RGB color value, scale is the anisotropic scale parameter, rotation is the rotation represented by a quaternion, and local density This represents the local point density feature. Further, the obtained feature vector is standardized: F norm =(F i -μ) / (σ+ε,F norm For the standardized eigenvalues, F i For each feature vector before standardization, μ is the mean of the corresponding feature value, σ is the standard deviation of the corresponding feature value, and ε is a preset coefficient value. The feature values ​​obtained after standardization can be combined to form the corresponding point cloud feature vector.

[0062] The basic scene model is clustered and segmented based on the established hierarchical clustering rules and the obtained point cloud feature vectors to obtain corresponding clusters. Specifically, the basic scene model can be coarsely segmented based on the position of the point cloud feature vectors within the basic scene model. The coarse segmentation process can be represented as: Cluster coarse =KMeans(F position ,k auto ), F position KMeans represents the position of the point cloud feature vector in the basic scene model. KMeans indicates that coarse segmentation uses the KMeans segmentation algorithm, which performs K-means clustering based on spatial location. auto This means that the value of k in the above coarse segmentation process can be adaptively determined, such as setting k... auto =min(20,N points / 100), N points Cluster represents the number of point cloud feature vectors. coarse This represents the coarse clusters obtained from coarse segmentation.

[0063] If the number of samples in a coarse cluster exceeds a preset threshold, the coarse cluster is further divided according to the fine clustering parameters to obtain the corresponding fine cluster (i.e., fine-grained segmentation results). The resulting fine cluster is then usable in subsequent steps. Specifically, if the number of samples in a coarse cluster exceeds a preset threshold, the DBSCAN algorithm is executed on the full feature data corresponding to that cluster for fine clustering. The DBSCAN parameters are set to a neighborhood radius (eps) of 0.3 and a minimum number of samples required to form a cluster (min...). samples If the number of samples in the coarse cluster is less than or equal to the threshold, then no further splitting is necessary, and the coarse cluster can be used as the cluster for subsequent steps.

[0064] Furthermore, the regions contained in the obtained clusters are merged and optimized according to the merging rules to obtain the corresponding key region set. Specifically, for each cluster forming a segmentation region R... k The center point μ can be calculated. k =mean(p i ∈Rk), bounding box: BB k =[min(p i ),max(p i Volume V k =∏(BB max -BB min Semantic consistency score (obtained based on the risk text information corresponding to the segmented region), point density ρ k =|Rk | / V k Based on the parameters of the segmented regions calculated in the above steps, the segmented regions are merged and optimized. Adjacent similar regions are merged based on the following criteria: spatial distance d(Ra,Rb) < preset distance threshold; feature similarity: sim(Fa,Fb) > preset similarity threshold; semantic consistency: having the same semantic label. After merging and optimizing the segmented regions, a set of key regions is obtained, which includes multiple segmented regions.

[0065] In a specific embodiment, obtaining the weighted risk coefficient corresponding to each segmented region in the key region set based on the point cloud feature vector and the feature description information in the feature information includes: performing semantic mapping on the feature description information in the feature information to obtain the corresponding semantic risk factor; performing aggregation analysis on the point cloud feature vector of the segmented region to obtain the geometric risk factor corresponding to each segmented region; determining the contextual risk factor corresponding to each segmented region based on the feature description information and the segmented regions adjacent to each segmented region; and performing weighted calculation on the semantic risk factor, the geometric risk factor, and the contextual risk factor according to a preset weighted risk model to obtain the weighted risk coefficient of each segmented region.

[0066] Furthermore, weighted risk coefficients are obtained for each segmented region in the key region set based on the feature vectors of the point cloud and the feature description information in the feature information. Specifically, semantic mapping can be performed on the feature description information corresponding to the segmented regions in the feature information to obtain the corresponding semantic risk factors; for example, the defined semantic risk mapping table is: {"road":0.9,"vehicles":0.8,"pedestrians":0.7,"machinery":0.85,"fire source":0.95,"furniture":0.5,"buildings":0.4,"vegetation":0.2,"terrain":0.1,"sky":0.05}.

[0067] The point cloud feature vectors contained in each segmented region are aggregated and analyzed to obtain the geometric risk factor corresponding to each segmented region. For example, the geometric risk factor r geom-k =g1×r volume +g2×r position +g3×r shape Among them, volume risk r volume =min(V k / V max Location risk r (1.0) position =1-min(d center / d max (1.0), shape risk r shape =min(aspect ratio / 10,1.0); g1, g2 and g3 are all weighting coefficients in the formula.

[0068] Based on the feature description information and the segmentation regions adjacent to each segmentation region, the feature description information of the segmentation regions adjacent to each segmentation region is obtained, and the risk factors corresponding to the feature description information of the adjacent segmentation regions are extracted as context risk factors; for example, the context risk factor is r. context-k =f(proximity risk, dynamism, functional importance), proximity risk: increases when near high-risk areas, dynamism: higher risk when near moving or variable objects, functional importance: higher risk when near key functional areas.

[0069] The weighted risk model calculates the weighted risk coefficient for each segmented region by weighting the semantic risk factor, geometric risk factor, and contextual risk factor. Specifically, the weighted risk model can be set as r. base-k =α b ×r geom-k +β b ×r sem-k +γ b ×r context-k α b β b and γ b These are all coefficient values ​​set in the weighted risk model, where α b +β b +γ b =1, r base-k This is the output of the weighted risk model. Normalizing this output yields the weighted risk coefficient r. base-k =clip(r base-k ,0.0,1.0).

[0070] Based on the basic risk weights of each segmented region, the weights of each segmented region in the basic scenario model are configured to obtain the corresponding hidden danger scenario model.

[0071] In a specific embodiment, after configuring the basic scenario model with weights according to the basic risk weights of each segmented region to obtain the corresponding hidden danger scenario model, the method further includes: optimizing the modeling density of the hidden danger scenario model according to a preset modeling density optimization strategy to obtain an optimized hidden danger scenario model.

[0072] Furthermore, after obtaining the hazard scenario model, the modeling density can be optimized using a modeling density optimization strategy to obtain an optimized hazard scenario model. Specifically, a risk-conditional 3DGS generation model is designed, whose loss function, based on the standard reconstruction loss Lrecon, introduces a risk distribution regularization term: ; where D(μ|R k ) represents region R k The spatial density of the inner Gaussian center μ is calculated using Var, which measures the variance of density between different regions. The first term encourages higher modeling detail density in high-risk regions, while the second term penalizes excessive differences in density between regions to prevent over-concentration.

[0073] In the process of optimizing modeling density, in addition to using multi-view image supervision to optimize appearance and geometry, the aforementioned risk regularization loss L is also used. risk-reg To guide the optimization of the initial distribution and density of Gaussian points in three-dimensional space, the generation process incorporates prior knowledge of risk. The Gaussian point optimization calculation process involves: targeting L... risk-reg The first term α·k∑D(μ|R) k To minimize the distribution deviation of Gaussian points within each region, typically D(μ|R) k This can be the KL divergence, Wasserstein distance, or negative log-likelihood. For L... risk-reg The second item - λ·‖r base - k‖2², maximizing the distance between the Gaussian point region and the baseline point (due to the negative sign), encourages the distribution of Gaussian points to move away from known risk areas. For L... risk-reg The third term β·Var({μ i Minimize the variance of Gaussian point locations within the region, and encourage the uniform distribution of Gaussian points within the region.

[0074] The specific optimization steps are as follows: First, perform five initialization settings: i1. Determine the three-dimensional spatial range [x min ,x max ]×[y min ,y max ]×[z min ,z max ], i2. Set the total number of Gaussian points N, i3. Divide into K regions R k (k=1,...,K), i4. Initialize the Gaussian point position μ i ^(0) (i=1,...,N), i5. Set the learning rate η and the maximum number of iterations T. After initialization, perform iterative optimization. The iterative steps are as follows: o1. Region allocation: allocate each Gaussian point μ i Assigned to the nearest region R k : k(i)=argmin k ||μ i -center(R k )‖;o2. Calculate each component of the loss function: L dist =α·∑ k ∑ {μi∈Rk} ||μi -μ Rk ||², L dist L represents the loss value corresponding to the distribution metric. base =-λ·∑ k ||r base -center(R k )‖²,L base L represents the loss value corresponding to the reference distance. var =β·∑ k Var({μ i |μ i ∈R k}), L var Indicates the loss value corresponding to the variance term; o3. Gradient differentiation: for each Gaussian point Its meaning is the gradient of the distribution term (corresponding to the distribution metric); +2λ·(r base -center(R k (i))), which means the gradient of the benchmark term (corresponding to the benchmark distance); +2β·(μ i -mean({μ j ∈R k (i)})), which means the gradient of the variance term (corresponding to the variance term). o4. Position update: ;o5. Density Adjustment: Calculate the regional weight w k ∝exp(-‖r base -center(R k After adjusting the region density, the optimized hazard scenario model is obtained. Specifically, for high-density regions, Gaussian points are increased or covariance is decreased; for low-density regions, Gaussian points are decreased or covariance is increased.

[0075] In the later stages of optimization or specific optimization phases, targeted "distribution equalization" iterations are activated; this process dynamically detects the Gaussian point in each region. R k The uniformity of the distribution of}, if the density in a certain area exceeds a threshold D max If it exerts a repulsive force on the Gaussian points at its edge, it can either guide the Gaussian points to diffuse repulsively into the adjacent low-density region.

[0076] Explanation of the repulsion-diffusion principle: (1) Definition and assumptions, Gaussian point set: Let the set of all Gaussian points be denoted as . G ={x i} i =1~ N , where x i ∈R2 (or R3) represents the location of a point; Region division: the space is divided into K one region { R k}k =1~ K Regional density: region R k At any moment t The density is defined as the number of Gaussian points in the region divided by its area (or volume): ρk ( t )=∣{x i ( t )∈ R k}∣ / Area( R k Density threshold: sets a maximum allowable density. D max (2) Dynamic detection and triggering conditions, at each update time t The system calculates the density of all regions. ρk ( t If a region exists R A Make: ρ A ( t )> D max This will trigger an action on the region. R A The "repulsion-diffusion" operation. (3) Apply repulsive force to the edge point, the edge point is defined as: in the region R A Within this, define its "edge point set". E A For those areas less than a certain distance from the region boundary δ The points, or more mathematically, using distance-based functions Judgment: ,in It is a region R A The boundary. The direction and target of the repulsive force: the target of the repulsive force is to push the point towards a neighboring low-density region. Let it be... R A The set of adjacent regions is { R B The form of force: for each edge point x i ∈ E A Calculate a pointer pointing to all adjacent low-density regions ( ρ B < ρ A The composite repulsive force vector of ( ). A simple definition is that for each such neighboring region R B The direction of the force is from the region RA center c A Pointing to area R B center c B Its size is related to the relative "openness" of the neighbor ( D max - ρ B It is directly proportional to and related to point x. i The force is inversely proportional to the distance to the boundary (the closer to the boundary, the greater the force). , among which, among which It is a small constant to prevent division by zero. (4) Diffusion process (position update): the edge points subjected to force update their positions according to the repulsive force, realizing diffusion to the neighboring region: x i ( t +1)=x i ( t )+ η· Frepel(x i );in η >0 is a coefficient that controls the diffusion step size (learning rate or time step). After the update, points may move from the region... R A Move to an adjacent area R B Inside.

[0077] S170. Reasoning is performed based on the aforementioned hidden danger scenario model to generate a corresponding spatial Gaussian scenario.

[0078] The input is a multi-view image not present in the initial image set or historical hazard image set. Forward reasoning is performed using a trained hazard scene model to generate a preliminary 3D Gaussian scene that implicitly contains risk distribution characteristics. The resulting 3D Gaussian scene is the spatial Gaussian scene.

[0079] S180. Based on the spatial scene characteristics of the spatial Gaussian scene, the basic risk weights of key areas in the hidden danger scene model are dynamically adjusted to generate a corresponding comprehensive risk weight set.

[0080] Based on the spatial scene characteristics of the obtained spatial Gaussian scene, the basic risk weights of key areas in the hidden danger scene model are dynamically adjusted to generate a comprehensive risk weight set corresponding to the key areas; the areas included in the key area set are the key areas.

[0081] In a specific embodiment, step S180 includes the following sub-steps: obtaining risk characteristic factors for each of the key regions based on the spatial scene characteristics; calculating the basic risk weights for each of the key regions based on a preset comprehensive risk weight function and the risk characteristic factors to obtain the comprehensive risk weights corresponding to each key region; and aggregating the comprehensive risk weights for each of the key regions to obtain a corresponding comprehensive risk weight set.

[0082] Specifically, risk characteristic factors for each key region are obtained based on the spatial scene characteristics. In the reconstructed spatial Gaussian scene, each key region can be obtained separately. R k The geometric complexity (e.g., curvature variation), semantic content (number of devices), and spatial topology (e.g., whether it is located at a passage intersection) are used as risk characteristic factors for each key region. Feature measurement information corresponding to each key region is obtained from the spatial scene features of the spatial Gaussian scene, thus yielding the following indicators for each risk region: G k Geometric complexity factor (normalized rate of change of curvature), S k Semantic content factor (device density), T k The spatial topological factor (degree of intersection) can be used to obtain the risk characteristic factors of each key area.

[0083] The basic risk weights of each key region are calculated based on the comprehensive risk weight function and risk characteristic factors, thereby obtaining the comprehensive risk weights of each key region. Specifically, the comprehensive risk weight function can be: ,in, b k Basic risk weights (preset). All are adjustment coefficients (set to 0.3, 0.4, and 0.3 respectively in the example).

[0084] For example, in a certain scenario, an underground pipeline inspection area is divided into three key areas: R1, R2, and R3. The basic data are as follows: R1: basic weight b1=0.6, curvature change rate 0.8 (G1=0.8), number of devices 3 (area 10㎡, density S1=0.3), no intersection (T1=0); R2: basic weight b2=0.7, curvature change rate 0.5 (G2=0.5), number of devices 5 (area 10㎡, density S2=0.5), three-way intersection (T2=0.6); R3: basic weight b3=0.8, curvature change rate 0.9 (G3=0.9), number of devices 8 (area 10㎡, density S3=0.8), four-way intersection (T3=1.0).

[0085] Calculation process (adjustment coefficients are taken as αk=0.3, βk=0.4, γk=0.3): R1 comprehensive risk weight: w1=0.6×(1+0.3×0.8+0.4×0.3+0.3×0)=0.816; R2 comprehensive risk weight: w2=0.7×(1+0.3×0.5+0.4×0.5+0.3×0.6)=1.071; R3 comprehensive risk weight: w3=0.8×(1+0.3×0.9+0.4×0.8+0.3×1.0)=1.512.

[0086] Case results analysis: R3 has the highest risk weight (1.512) due to its high curvature variation, high equipment density, and four-way intersection; R2 has the second highest risk weight (1.071) due to its moderate equipment density and three-way intersection; R1 has the lowest risk weight (0.816) due to its lack of intersection and relatively few equipment.

[0087] By dynamically adjusting the basic risk weights of each key area, a final comprehensive risk weight set is generated. The comprehensive risk weight set then includes the comprehensive risk weights of each key region.

[0088] Using the calculated final risk weight map W The risk factor is used to perform non-rigid fine-tuning on the spatial Gaussian scene generated in step S170. By optimizing an energy function, the distribution of details (such as small objects and surface textures) in high-risk weight regions is made more consistent with their risk level while maintaining visual realism, and the overall distribution of all regions is ensured to meet the preset uniformity constraint.

[0089] S190. Match the spatial Gaussian scene with the hidden danger scene model according to the preset spatial grid matching algorithm and the comprehensive risk weight set to obtain the corresponding target hidden danger scene model.

[0090] Furthermore, based on the spatial grid matching algorithm and the comprehensive risk weight set obtained from the above steps, the spatial Gaussian scene is matched with the hazard scene model to obtain the corresponding target hazard scene model. The main construction idea for matching the spatial Gaussian scene with the hazard scene model is as follows: First, the spatial Gaussian scene is enclosed with an AABB box. Then, the large AABB box is cut into 10cm cubic grids (i.e., scene squares, implemented based on the spatial grid matching algorithm). Each scene square is assigned a risk weight assessment. At the same time, the hazard scene model is also enclosed with an AABB box. An octree is used to match the scene squares. Then, all suitable scene squares are added to a list. Finally, a sorting algorithm is used to find the optimal generation position.

[0091] In a specific embodiment, step S190 includes the following sub-steps: applying the spatial grid in the spatial grid matching algorithm to cut the spatial Gaussian scene to obtain corresponding scene squares; scoring each scene square according to the comprehensive risk weight set to obtain a score label for each scene square; judging whether the hazard model square in the hazard scene model intersects and collides with the scene square to obtain a corresponding collision judgment result; filtering the scene squares with no collision judgment result according to the square filtering rules in the spatial grid matching algorithm and the score label to obtain target scene squares; and placing the hazard model square that matches the target scene square in each of the target scene squares to generate the corresponding target hazard scene model.

[0092] First, the spatial Gaussian scene is segmented using the spatial grid matching algorithm to form scene squares. The specific steps include: traversing all vertices of the spatial Gaussian scene from the previous step to obtain the minimum and maximum values ​​on the x, y, and z axes; using the minimum and maximum values ​​of each axis as the boundaries of the bounding box to obtain the AABB box; and, according to the size of the spatial grid (e.g., set to 10×10×10cm), cutting the AABB box into multiple scene squares by cutting along the three axes of the AABB box in units of 10cm to obtain a finite number of scene squares.

[0093] Next, each scenario grid is scored based on the comprehensive risk weight set, resulting in a score label for each scenario grid. The scoring rules are as follows: based on the scenario model's generation rules and the comprehensive risk weight set, each key area is divided into high-risk, medium-risk, and normal areas. The score changes from 100 to 0, corresponding to a progression from high-risk to normal.

[0094] Furthermore, the system determines whether the hazard model squares in the hazard scene model intersect or collide with scene squares, obtaining a collision judgment result. That is, the score of each scene square within the current critical area is marked as the score for that critical area. This process requires determining whether a scene square is located within a critical area, using a collision algorithm between two objects. If a scene square intersects or collides with a critical area of ​​the hazard scene model, the score for that critical area is directly marked as -1, indicating that the critical area cannot be placed there.

[0095] Furthermore, based on the grid selection rules and scoring tags in the spatial grid matching algorithm, scene grids with non-collision results are filtered to obtain the target scene grids. Specifically, all 10cm grids in the scene are traversed. 3 Scene grids, and scene grids in the collision judgment results where the hazard scene model did not collide with a scene grid with a negative score, can all be placed in the grid list.

[0096] The scene squares placed in the grid list are filtered according to the following rules: Sort the scene squares in the grid list according to their scores and keep all scene squares with the highest scores; Sort the scene squares retained in the previous step according to their distance from the nearest high-risk area and keep all scene squares that are closest to the high-risk area; Compare the scene squares retained in the previous step with the distances of all existing hazard models in the spatial Gaussian scene and keep the scene square with the farthest distance as the final target scene square.

[0097] Hazard model squares that match the target scene squares are placed in each target scene square to generate the corresponding target hazard scene model. The hazard model squares are obtained by dividing the hazard scene model through a spatial grid. There is a positional correspondence between the target scene squares and the hazard model squares. Therefore, the hazard model squares that are in the same position as the target scene squares are the hazard model squares that match the target scene squares.

[0098] The hazard model generation method based on intelligent recognition and spatial grid algorithms disclosed in the above embodiments includes: preprocessing and differentially labeling an initial image set to obtain a structured labeled dataset, then training a corresponding hazard recognition neural network; identifying historical hazard image sets using the hazard recognition neural network to obtain hazard recognition results and target risk intervals for hazard areas; constructing a corresponding hazard scene model and inferring a spatial Gaussian scene; dynamically adjusting the basic risk weights of key areas in the hazard scene model and matching the spatial Gaussian scene with the hazard scene model to obtain a corresponding target hazard scene model. This hazard model generation method automatically generates a simulation scene model corresponding to real-world hazards through virtual simulation and 3D interactive technology. Users can intuitively learn hazard characteristics and practice / assess in a virtual environment, improving the efficiency of user training and enhancing the immersion and practicality of the training.

[0099] This invention also provides a hazard model generation device based on intelligent identification and spatial grid algorithms. This device can be configured in a terminal device and is used to execute any embodiment of the aforementioned hazard model generation method based on intelligent identification and spatial grid algorithms. Specifically, please refer to... Figure 2 , Figure 2 This is a schematic block diagram of a hazard model generation device based on intelligent identification and spatial grid algorithms provided in an embodiment of the present invention.

[0100] like Figure 2As shown, the hazard model generation device 100 based on intelligent recognition and spatial grid algorithm includes an image preprocessing unit 110, a labeling unit 120, a network training unit 130, a recognition unit 140, a risk analysis unit 150, a scene model construction unit 160, a scene generation unit 170, a weight adjustment unit 180, and a target hazard scene model acquisition unit 190.

[0101] Image preprocessing unit 110 is used to receive the input initial image set and preprocess the images contained in the initial image set to obtain a corresponding preprocessed image set.

[0102] The annotation unit 120 is used to perform differential annotation on the images contained in the preprocessed image set according to the preset differential annotation rules, so as to obtain the corresponding structured annotation dataset.

[0103] The network training unit 130 is used to train the pre-stored initial identification network based on the structured labeled dataset to obtain the corresponding hidden danger identification neural network.

[0104] The identification unit 140 is used to receive the input set of historical hazard images, and to identify the set of historical hazard images according to the hazard identification neural network to obtain the corresponding hazard identification result.

[0105] The risk analysis unit 150 is used to perform risk analysis on each hidden danger area in the historical hidden danger image set according to the preset risk probability analysis rules and the hidden danger identification results, so as to obtain the target risk interval corresponding to each hidden danger area.

[0106] The scenario model construction unit 160 is used to construct a corresponding hazard scenario model based on the feature information corresponding to the historical hazard image set, the target risk range, and the risk probability analysis information.

[0107] The scene generation unit 170 is used to perform reasoning based on the hidden danger scene model to generate a corresponding spatial Gaussian scene.

[0108] The weight adjustment unit 180 is used to dynamically adjust the basic risk weights of key areas in the hidden danger scenario model according to the spatial scene characteristics of the spatial Gaussian scene, so as to generate a corresponding comprehensive risk weight set.

[0109] The target hazard scene model acquisition unit 190 is used to match the spatial Gaussian scene with the hazard scene model according to the preset spatial grid matching algorithm and the comprehensive risk weight set, so as to obtain the corresponding target hazard scene model.

[0110] The hazard model generation device based on intelligent recognition and spatial grid algorithms provided in this embodiment of the invention applies the aforementioned hazard model generation method based on intelligent recognition and spatial grid algorithms. After preprocessing and differentially labeling the initial image set to obtain a structured labeled dataset, a corresponding hazard recognition neural network is trained. Based on the hazard recognition neural network, a historical hazard image set is identified to obtain hazard recognition results and the target risk interval of the hazard area is obtained. A corresponding hazard scene model is constructed, and a spatial Gaussian scene is inferred. The basic risk weights of key areas in the hazard scene model are dynamically adjusted, and the spatial Gaussian scene is matched with the hazard scene model to obtain the corresponding target hazard scene model. This hazard model generation method automatically generates a simulation scene model corresponding to real-world hazards through virtual simulation and 3D interactive technology. Users can intuitively learn hazard characteristics and practice / assess in a virtual environment, improving the efficiency of user training and learning, while enhancing the immersion and practicality of the training.

[0111] The aforementioned hazard model generation device based on intelligent identification and spatial grid algorithms can be implemented as a computer program, which can be used in various ways, such as... Figure 3 It runs on the computer device shown.

[0112] Please see Figure 3 , Figure 3 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. The computer device can be a terminal device used to execute a hazard model generation method based on intelligent recognition and spatial grid algorithms to generate a corresponding target hazard scene model for training and learning.

[0113] See Figure 3 The computer device 500 includes a processor 502, a memory, and a communication interface 505 connected via a communication bus 501. The memory may include a storage medium 503 and internal memory 504.

[0114] The storage medium 503 can store the operating system 5031 and the computer program 5032. When the computer program 5032 is executed, it enables the processor 502 to execute a hazard model generation method based on intelligent identification and spatial grid algorithms. The storage medium 503 can be a volatile storage medium or a non-volatile storage medium.

[0115] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0116] The internal memory 504 provides an environment for the computer program 5032 in the storage medium 503 to run. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a method for generating a hazard model based on intelligent identification and spatial grid algorithms.

[0117] This communication interface 505 is used for network communication, such as providing data transmission. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device 500 to which the present invention is applied. The specific computer device 500 may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0118] The processor 502 is used to run the computer program 5032 stored in the memory to realize the corresponding functions in the above-mentioned method for generating hazard models based on intelligent identification and spatial grid algorithms.

[0119] Those skilled in the art will understand that Figure 3 The embodiments of the computer device shown do not constitute a limitation on the specific configuration of the computer device. In other embodiments, the computer device may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. For example, in some embodiments, the computer device may include only memory and a processor. In such embodiments, the structure and function of the memory and processor are different from those shown. Figure 3 The embodiments shown are consistent and will not be described again here.

[0120] It should be understood that, in this embodiment of the invention, the processor 502 may be a Central Processing Unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), microcontroller units (MCUs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0121] In another embodiment of the invention, a computer-readable storage medium is provided. This computer-readable storage medium may be volatile or non-volatile. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps included in the aforementioned method for generating a hazard model based on intelligent identification and spatial grid algorithms.

[0122] Those skilled in the art will readily understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

[0123] In the embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Units with the same function may be grouped into one unit. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, or it may be an electrical, mechanical, or other form of connection.

[0124] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention, depending on actual needs.

[0125] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0126] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a computer-readable storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned computer-readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks.

[0127] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for generating hazard models based on intelligent identification and spatial grid algorithms, characterized in that, The method includes: Receive the input initial image set, and preprocess the images contained in the initial image set to obtain the corresponding preprocessed image set; The images in the preprocessed image set are differentially labeled according to the preset differential labeling rules to obtain the corresponding structured labeled dataset; The pre-stored initial identification network is trained based on the structured labeled dataset to obtain the corresponding hazard identification neural network; The system receives a set of historical hazard images and identifies them using the hazard identification neural network to obtain the corresponding hazard identification results. Based on the preset risk probability analysis rules and the hazard identification results, risk analysis is performed on each hazard area in the historical hazard image set to obtain the target risk interval corresponding to each hazard area. Based on the feature information corresponding to the historical hazard image set, the target risk range, and the risk probability analysis information, a corresponding hazard scenario model is constructed. Reasoning is performed based on the aforementioned potential hazard scenario model to generate a corresponding spatial Gaussian scenario; Based on the spatial scene characteristics of the Gaussian spatial scene, the basic risk weights of key areas in the hidden danger scene model are dynamically adjusted to generate a corresponding comprehensive risk weight set. The spatial Gaussian scene and the hidden danger scene model are matched according to the preset spatial grid matching algorithm and the comprehensive risk weight set to obtain the corresponding target hidden danger scene model.

2. The method for generating a hazard model based on intelligent identification and spatial grid algorithm according to claim 1, characterized in that, The step of performing differential annotation on the images in the preprocessed image set according to preset differential annotation rules to obtain the corresponding structured annotation dataset includes: The images in the preprocessed image set are classified according to the classification types included in the differential annotation rules to obtain the classification image sets corresponding to each classification type; According to the point-shaped hazard annotation format in the differentiated annotation rules, the classification image set corresponding to the point-shaped hazard type is annotated to obtain the corresponding point-shaped hazard annotation information; According to the rule hazard annotation format in the differentiated annotation rules, the classification image set corresponding to the rule hazard type is annotated to obtain the corresponding rule hazard annotation information; According to the irregular hazard labeling format in the differentiated labeling rules, the classification image set corresponding to the irregular hazard type is labeled to obtain the corresponding irregular hazard labeling information; The point-like hazard labeling information, the regular hazard labeling information, and the irregular hazard labeling information are integrated to obtain the corresponding structured labeling dataset.

3. The method for generating a hazard model based on intelligent identification and spatial grid algorithms according to claim 1, characterized in that, The step involves performing risk analysis on each hazard area in the historical hazard image set based on preset risk probability analysis rules and the hazard identification results, to obtain the target risk interval corresponding to each hazard area, including: Based on the historical hazard image set, the empirical interval of each hazard area is obtained by matching the risk weight range in the risk probability analysis rule; Obtain the hazard information corresponding to each hazard area from the hazard identification results; The corresponding risk statistical index is calculated based on the risk value calculation formula in the risk probability analysis rules for each of the hidden danger areas; The risk statistical index is mapped to the empirical interval of the corresponding hidden danger area to obtain the interval mapping value corresponding to each hidden danger area as the risk probability analysis information; The empirical intervals are corrected based on the observation data corresponding to the historical hazard image set and the risk probability analysis information to obtain the updated target risk intervals for each hazard area.

4. The method for generating a hazard model based on intelligent identification and spatial grid algorithm according to any one of claims 1-3, characterized in that, The step of constructing a corresponding hazard scenario model based on the feature information corresponding to the historical hazard image set, the target risk range, and the risk probability analysis information includes: Virtual modeling is performed based on the image location information corresponding to each image in the historical hidden danger image set, in order to obtain the corresponding basic scene model; The corresponding point cloud feature vectors are extracted from each image in the historical potential hazard image set; The basic scene model is clustered and segmented according to the preset hierarchical clustering segmentation rules and the point cloud feature vectors to obtain the corresponding clusters; The regions contained in the clusters are merged and optimized according to the preset merging rules to obtain the corresponding set of key regions; Based on the point cloud feature vector and the feature description information in the feature information, obtain the weighted risk coefficient corresponding to each segmented region in the key region set; The weighted risk coefficients of each segmented region are normalized according to the target risk interval to obtain the basic risk weights corresponding to each segmented region. The basic scenario model is weighted according to the basic risk weight of each segmented region to obtain the corresponding hidden danger scenario model.

5. The method for generating a hazard model based on intelligent identification and spatial grid algorithm according to claim 4, characterized in that, After configuring the basic scenario model with weights based on the basic risk weights of each segmented region to obtain the corresponding hidden danger scenario model, the method further includes: The modeling density of the hazard scene model is optimized according to the preset modeling density optimization strategy to obtain the optimized hazard scene model.

6. The method for generating a hazard model based on intelligent identification and spatial grid algorithms according to claim 5, characterized in that, The step of dynamically adjusting the basic risk weights of key areas in the hazard scenario model based on the spatial scenario characteristics of the hazard scenario model to generate a corresponding comprehensive risk weight set includes: Risk characteristic factors for each of the key regions are obtained based on the spatial scene characteristics; The basic risk weights of each key region are calculated based on the preset comprehensive risk weight function and the risk characteristic factors to obtain the comprehensive risk weights corresponding to each key region. The comprehensive risk weights of each key region are aggregated to obtain the corresponding comprehensive risk weight set.

7. The method for generating a hazard model based on intelligent identification and spatial grid algorithms according to claim 5, characterized in that, The step of matching the spatial Gaussian scene with the hidden danger scene model according to the preset spatial grid matching algorithm and the comprehensive risk weight set to obtain the corresponding target hidden danger scene model includes: The spatial Gaussian scene is segmented using the spatial grid matching algorithm to obtain the corresponding scene grid. Each scenario grid is scored based on the comprehensive risk weight set to obtain a score label for each scenario grid; The system determines whether the hazard model grid in the hazard scene model intersects or collides with the scene grid, and obtains the corresponding collision judgment result. The scene squares with no collision judgment result are filtered according to the square filtering rules in the spatial grid matching algorithm and the scoring mark to obtain the target scene squares. Place the hazard model grid that matches the target scene grid into each of the target scene grids to generate the corresponding target hazard scene model.

8. A hazard model generation device based on intelligent identification and spatial grid algorithms, characterized in that, The apparatus is used to execute the hazard model generation method based on intelligent identification and spatial grid algorithm as described in any one of claims 1-7, the apparatus comprising: The image preprocessing unit is used to receive the input initial image set and preprocess the images contained in the initial image set to obtain the corresponding preprocessed image set. The annotation unit is used to perform differential annotation on the images contained in the preprocessed image set according to the preset differential annotation rules to obtain the corresponding structured annotation dataset; The network training unit is used to train the pre-stored initial identification network based on the structured labeled dataset to obtain the corresponding hidden danger identification neural network. The identification unit is used to receive the input set of historical hazard images, and to identify the set of historical hazard images according to the hazard identification neural network to obtain the corresponding hazard identification result; The risk analysis unit is used to perform risk analysis on each hidden danger area in the historical hidden danger image set according to the preset risk probability analysis rules and the hidden danger identification results, and to obtain the target risk interval corresponding to each hidden danger area. The scenario model construction unit is used to construct a corresponding hazard scenario model based on the feature information corresponding to the historical hazard image set, the target risk range, and the risk probability analysis information. The scene generation unit is used to perform reasoning based on the hidden danger scene model to generate a corresponding spatial Gaussian scene; The weight adjustment unit is used to dynamically adjust the basic risk weights of key areas in the hidden danger scenario model according to the spatial scene characteristics of the spatial Gaussian scene, so as to generate a corresponding comprehensive risk weight set. The target hazard scene model acquisition unit is used to match the spatial Gaussian scene with the hazard scene model according to the preset spatial grid matching algorithm and the comprehensive risk weight set, so as to obtain the corresponding target hazard scene model.

9. A computer device, characterized in that, The device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements the steps of the method for generating a hazard model based on intelligent identification and spatial grid algorithms as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for generating a hazard model based on intelligent identification and spatial grid algorithm as described in any one of claims 1-7.

Images