Environment modeling method and system for realizing intelligent security and protection

Abstract

The invention discloses an environment modeling method and system for realizing intelligent security and protection, and relates to the technical field of intelligent security and protection, and the method comprises the steps: collecting a visual image and depth data of a monitoring region, and generating a monitoring scene feature matrix; performing hierarchical feature extraction on the monitoring scene feature matrix by using a deep convolutional neural network, and constructing a scene semantic feature map; projecting the scene semantic feature map to a three-dimensional monitoring space according to the depth data, and constructing a three-dimensional scene grid model; extracting spatial topological characteristics of each grid node in the three-dimensional scene grid model, and calculating association weights among the grid nodes; and constructing a scene dynamic knowledge graph based on the spatial topological features and the associated weights, and carrying out modeling analysis on abnormal events. According to the method, through cooperative collection of the multi-view vision sensor and the depth sensor, in combination with information entropy weighted standardization and principal component analysis, accurate generation of the monitoring scene feature matrix is realized, and the problems of information redundancy and feature coupling in traditional vision data processing are effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent security technology, and in particular to an environment modeling method and system for realizing intelligent security. Background Art

[0002] Traditional security monitoring systems mainly rely on fixed-viewing cameras and simple motion detection algorithms, which severely limit their perception and scene understanding capabilities. These systems can usually only capture two-dimensional images and cannot accurately understand the spatial structure and semantic information of the monitoring scene, resulting in low accuracy and real-time detection of abnormal events. Especially in complex and dynamically changing environments, traditional methods often have difficulty in effectively identifying potential security risks.

[0003] With the rapid development of artificial intelligence and computer vision technology, security systems are facing the need to transform from passive monitoring to active perception and intelligent analysis. Existing technologies still have key technical bottlenecks such as inaccurate feature extraction, in-depth scene semantic understanding, and lack of real-time modeling of dynamic changes in the environment. Summary of the invention

[0004] In view of the limitations of existing security technologies, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is how to improve the perception capability and abnormal event detection accuracy of the security system through multi-dimensional feature extraction, semantic understanding and knowledge graph construction.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, an embodiment of the present invention provides an environmental modeling method for realizing intelligent security, which includes using multi-eye vision sensors and depth sensors to collect visual images and depth data of the monitoring area to generate a monitoring scene feature matrix; using a deep convolutional neural network to perform hierarchical feature extraction on the monitoring scene feature matrix, and introducing a spatial-channel dual attention module to calibrate important area and channel features to construct a scene semantic feature map; projecting the scene semantic feature map into a three-dimensional monitoring space according to the depth data to construct a three-dimensional scene grid model; extracting the spatial topological features of each grid node in the three-dimensional scene grid model, and calculating the association weights between the grid nodes; constructing a scene dynamic knowledge graph based on the spatial topological features and the association weights, and modeling and analyzing abnormal events.

[0008] As a preferred solution of the environmental modeling method for realizing intelligent security described in the present invention, generating a monitoring scene feature matrix includes the following steps: performing image block division on the visual image, extracting the color distribution histogram and texture feature vector of each image block; performing region segmentation on the depth data, calculating the depth gradient histogram and curvature feature vector of each region; performing information entropy weighted standardization preprocessing on the color distribution histogram, texture feature vector, depth gradient histogram, and curvature feature vector to obtain a standardized feature fusion matrix; performing dimensionality reduction and feature decoupling on the standardized feature fusion matrix through principal component analysis to extract the main information components; and reconstructing the main information components into a monitoring scene feature matrix.

[0009] As a preferred solution of the environmental modeling method for realizing intelligent security described in the present invention, wherein: constructing a scene semantic feature map includes the following steps: constructing a multi-scale convolutional hierarchical network, and performing feature decomposition on the monitoring scene feature matrix through convolution kernels of different scales; establishing a spatial attention module, calculating the spatial position correlation score after feature decomposition, and generating a spatial weight matrix; establishing a channel attention module, calculating the channel correlation score after feature decomposition, and generating a channel weight matrix; mapping the spatial weight matrix and the channel weight matrix to the feature space respectively, and obtaining a dual attention weight map through nonlinear transformation; dynamically enhancing the feature decomposition results according to the dual attention weight map to generate a scene semantic feature map.

[0010] As a preferred solution of the environmental modeling method for realizing intelligent security described in the present invention, the following steps are included in the construction of a three-dimensional scene grid model: using depth data to perform three-dimensional coordinate back projection on feature points in a scene semantic feature map to generate a feature point cloud; preprocessing the feature point cloud; performing spatial clustering and segmentation on the preprocessed feature point cloud to establish a local feature area; constructing an octree spatial index based on the local feature area, dividing the three-dimensional grid units, and adaptively adjusting the grid resolution and octree depth according to the complexity of the scene; in each three-dimensional grid unit, reconstructing a three-dimensional surface grid using a Poisson reconstruction algorithm; splicing the three-dimensional surface grids of adjacent three-dimensional grid units along the boundary to construct a three-dimensional scene grid model.

[0011] As a preferred solution of the environmental modeling method for realizing intelligent security described in the present invention, wherein: calculating the association weights between grid nodes includes the following steps: constructing a K nearest neighbor graph for the grid nodes in the three-dimensional scene grid model, and recording the local connection relationship of each grid node; calculating the geometric descriptor of each grid node, the geometric descriptor includes the normal vector, principal curvature and local surface shape coefficient of the grid node; based on the local connection relationship, using a graph convolutional neural network to perform feature aggregation on the grid nodes, and generate a high-order topological feature representation of the grid nodes; establishing a feature similarity matrix of the grid nodes, and calculating the spatial correlation in combination with the Euclidean distance between the nodes; according to the high-order topological feature representation and the spatial correlation, using the attention mechanism to calculate the association weights between the grid nodes; wherein the spatial topological features include the local connection relationship, geometric descriptor, high-order topological feature representation and spatial correlation of the grid nodes.

[0012] As a preferred solution of the environmental modeling method for realizing intelligent security described in the present invention, wherein: constructing a scene dynamic knowledge graph includes the following steps: defining node attribute vectors according to spatial topological features, constructing edge attribute vectors based on association weights, and forming basic elements of the scene knowledge graph; using a graph database to store the basic elements of the scene knowledge graph, and establishing a time series index mapping table; using an attention graph neural network to dynamically evolve the scene knowledge graph and generate a scene state embedding representation; designing a multi-task learning module, and inputting the scene state embedding representation into the anomaly classifier and the positioning regressor respectively; fusing the output results of the anomaly classifier and the positioning regressor to generate a spatiotemporal feature description of the abnormal event.

[0013] As a preferred solution of the environmental modeling method for realizing intelligent security described in the present invention, wherein: feature decomposition of the monitoring scene feature matrix through convolution kernels of different scales includes the following steps: constructing a multi-scale convolution hierarchical network, the multi-scale convolution hierarchical network includes parallel convolution branches, and the convolution kernel sizes of the parallel convolution branches are configured according to an increasing rule; inputting the monitoring scene feature matrix into the parallel convolution branch, and obtaining convolution features through convolution kernels of different scales; setting a residual connection unit in the parallel convolution branch, and the residual connection unit combines the input features with the convolution features; performing channel splicing on the features output by the parallel convolution branch to obtain multi-scale fusion features; using a singular value decomposition algorithm to process the multi-scale fusion features to generate feature decomposition results.

[0014] In the second aspect, an embodiment of the present invention provides an environmental modeling system for realizing intelligent security, which includes a data acquisition module, which is used to use multi-eye vision sensors and depth sensors to collect visual images and depth data of the monitoring area, and generate a monitoring scene feature matrix; a feature extraction module, which is used to use a deep convolutional neural network to perform hierarchical feature extraction on the monitoring scene feature matrix, and introduce a spatial-channel dual attention module to calibrate important area and channel features, and construct a scene semantic feature map; a three-dimensional modeling module, which is used to project the scene semantic feature map into a three-dimensional monitoring space according to the depth data, and construct a three-dimensional scene grid model; a topological feature extraction module, which is used to extract the spatial topological features of each grid node in the three-dimensional scene grid model, and calculate the association weights between the grid nodes; an abnormal module, which is used to construct a scene dynamic knowledge graph based on the spatial topological features and the association weights, and model and analyze abnormal events.

[0015] The beneficial effects of the present invention are as follows: the present invention realizes the accurate generation of the monitoring scene feature matrix through the collaborative collection of multi-eye vision sensors and depth sensors, combined with information entropy weighted normalization and principal component analysis, and effectively solves the information redundancy and feature coupling problems in traditional visual data processing; introduces the space-channel dual attention module, and realizes the adaptive extraction and enhancement of scene semantic features by constructing a multi-scale convolutional hierarchical network and residual connection units, which significantly improves the robustness and discrimination of feature expression; constructs a three-dimensional scene grid model based on depth data, extracts grid node topological features, and constructs a dynamic knowledge graph based on an attention graph neural network, innovatively expanding scene understanding from a two-dimensional plane to a three-dimensional space, providing a more comprehensive and in-depth technical path for the accurate modeling and intelligent analysis of abnormal events. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 A flowchart of the framework of the environment modeling method for realizing intelligent security.

[0018] Figure 2 The process of constructing a three-dimensional scene mesh model for an environmental modeling method for realizing intelligent security. DETAILED DESCRIPTION

[0019] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0020] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0022] Example 1, reference Figure 1-2 , which is the first embodiment of the present invention, and provides an environment modeling method for realizing intelligent security, and the framework flow chart is as follows Figure 1 As shown, including,

[0023] S1: Use multi-view vision sensors and depth sensors to collect visual images and depth data of the monitoring area and generate a monitoring scene feature matrix.

[0024] Among them, the multi-eye vision sensor is arranged with a 120° field of view, and the interval between adjacent sensors ensures a 50% overlapping field of view to ensure stable feature matching. The depth sensor uses structured light technology, with a measurement accuracy of ±2mm and a sampling frequency of 60Hz, realizing all-round and high-precision perception of the monitoring scene.

[0025] In the process of generating the monitoring scene feature matrix, the visual image is divided into image blocks by using the overlapping sliding window method, and the visual image is divided into 32×32 pixel image blocks with an overlapping rate of 50% for adjacent blocks, and the HSV color distribution histogram and texture feature vector of each image block are extracted; the depth data is segmented based on the adaptive region growing algorithm, and the depth gradient histogram and curvature feature vector of each region are calculated; the HSV color distribution histogram, texture feature vector, depth gradient histogram and curvature feature vector are preprocessed by information entropy weighted standardization to obtain a standardized feature fusion matrix; the standardized feature fusion matrix is ​​reduced in dimension and decoupled by principal component analysis, and the principal components with cumulative variance contribution rate exceeding 90% are selected to extract the main information components; the main information components are reconstructed into the monitoring scene feature matrix.

[0026] Among them, the information entropy weighted standardization preprocessing process is as follows: first, the information entropy of each feature vector is calculated, and the feature weight is determined based on the size of the information entropy; then, the Z-score standardization method is used to normalize the scale of each feature to eliminate the influence of outliers and dimensional differences; finally, the standardized features are weighted and fused according to the weight coefficient to generate a standardized feature fusion matrix.

[0027] Preferably, for image block division, 32×32 pixels are selected as the basic unit and a 50% overlap rate is adopted, which not only ensures the local description ability of the feature, but also improves the robustness of the feature; for depth data region segmentation, an adaptive threshold strategy is used to dynamically adjust the depth gradient threshold and the connected domain area threshold to improve the segmentation effect; in principal component analysis, the principal component with a variance contribution rate of 90% is selected to avoid excessive information loss while ensuring the dimensionality reduction effect.

[0028] S2: A deep convolutional neural network is used to perform hierarchical feature extraction on the monitoring scene feature matrix, and a spatial-channel dual attention module is introduced to calibrate important area and channel features to construct a scene semantic feature map.

[0029] Specifically, the following steps are included:

[0030] S2.1: Construct a multi-scale convolutional hierarchical network and perform feature decomposition on the monitoring scene feature matrix through convolution kernels of different scales.

[0031] Specifically, a multi-scale convolutional hierarchical network is constructed, which includes parallel convolution branches, and the convolution kernel sizes of the parallel convolution branches are configured according to an increasing rule; the monitoring scene feature matrix is ​​input into the parallel convolution branch, and the convolution features are obtained by processing with convolution kernels of different scales; a residual connection unit is set in the parallel convolution branch, and the residual connection unit combines the input features with the convolution features to perform channel splicing on the features output by the parallel convolution branch to obtain multi-scale fusion features; the multi-scale fusion features are processed using a singular value decomposition algorithm to generate feature decomposition results.

[0032] Preferably, the parallel convolution branch includes three parallel branches, which respectively adopt convolution kernel sizes of 3×3, 5×5 and 7×7, each branch contains three convolution layers, the number of convolution kernels in the first layer is 64, the second layer is 128, and the third layer is 256, and the activation function adopts ReLU; the residual connection unit adopts a 1×1 convolution layer to achieve dimensionality reduction and feature matching, reduces the input feature to the same number of channels as the convolution feature, and then realizes feature combination by element-by-element addition; the output features of the three branches are spliced ​​in the channel dimension to obtain a multi-scale fusion feature of 768 channels; singular value decomposition is applied to the feature, and based on the threshold criterion that the feature reconstruction error is less than 0.1, the feature vectors corresponding to the first 384 largest singular values ​​are retained to achieve feature dimensionality reduction and decoupling.

[0033] S2.2: Establish a spatial attention module, calculate the spatial position correlation score after feature decomposition, and generate a spatial weight matrix.

[0034] Specifically, the spatial attention module adopts a dual-path feature extraction structure. First, the input features are max-pooled and average-pooled in the channel dimension to obtain two two-dimensional spatial feature maps. After the two feature maps are spliced ​​in the channel dimension, a 5×5 convolution layer is used to extract spatial correlation features, and the ReLU activation function is used to increase the nonlinear expression ability. Then a 1×1 convolution layer is used to compress the channel dimension, and finally the Sigmoid activation function is used to map the features to the [0,1] interval to generate a spatial attention weight map. The value of each position in the weight map represents the importance of the corresponding spatial position. The larger the value, the more critical the information contained in the position.

[0035] S2.3: Establish a channel attention module, calculate the channel correlation score after feature decomposition, and generate a channel weight matrix.

[0036] In the specific implementation, the channel attention module first performs global average pooling on the input features to obtain the channel descriptor; then, the channel dependency is constructed through two fully connected layers. The first fully connected layer reduces the number of channels to 1 / 4 of the original number and uses ReLU activation. The second fully connected layer restores the number of channels and uses Sigmoid activation to obtain the initial channel weights. In order to enhance the modeling of the relationship between channels, the correlation coefficient matrix between channel features is calculated, which is obtained by flattening the feature map, performing matrix inner product operations and normalizing it. Finally, the initial channel weights are fused with the correlation coefficient matrix by weighted averaging, and the weight coefficients are adaptively learned through back-propagation to generate the final channel weight matrix.

[0037] S2.4: Map the spatial weight matrix and the channel weight matrix to the feature space respectively, and obtain the dual attention weight map through nonlinear transformation.

[0038] Specifically, the spatial weight matrix is ​​first expanded to the same channel dimension as the original feature through a 1×1 convolution layer; at the same time, the channel weight matrix is ​​copied to the same spatial dimension as the original feature through a dimension expansion operation; the two weight feature maps are element-wise multiplied to obtain the initial fused feature; then the fused feature is processed through a feature reconstruction network consisting of two layers of 1×1 convolution. The first layer reduces the number of channels to 1 / 2 of the original feature channels and uses ReLU activation, and the second layer restores the original number of channels and uses Sigmoid activation, outputting a dual attention weight map to achieve effective fusion of attention information in the spatial dimension and channel dimension.

[0039] S2.5: Dynamically enhance the feature decomposition results according to the dual attention weight map to generate a scene semantic feature map.

[0040] Specifically, the dual attention weight map is element-wise multiplied with the feature decomposition result to obtain the attention-enhanced feature; the original feature decomposition result is added to the enhanced feature using a residual structure to ensure the retention of important information; the final scene semantic feature map is obtained through a feature reconstruction module, which contains two 3×3 convolutional layers, the first layer uses ReLU activation, the second layer uses ReLU activation, and the number of output channels is the same as the original feature. In order to enhance the multi-scale expression ability of the feature, two branches are introduced in the feature reconstruction process: the main branch maintains the original resolution, and the auxiliary branch captures information with a larger receptive field through 2x downsampling-upsampling. Finally, the feature maps of the two branches are fused in a weighted manner, and the weight coefficient is adaptively determined by trainable parameters.

[0041] S3: Project the scene semantic feature map into the three-dimensional monitoring space based on the depth data and construct a three-dimensional scene mesh model.

[0042] Specifically, the process of constructing the 3D scene mesh model is as follows: Figure 2 As shown, the following steps are included:

[0043] S3.1: Use depth data to perform three-dimensional coordinate back-projection on the feature points in the scene semantic feature map to generate a feature point cloud.

[0044] Specifically, the camera intrinsic parameter matrix and extrinsic parameter matrix are used for coordinate transformation to convert each pixel in the depth image into a three-dimensional point in the world coordinate system; the feature points with a confidence level greater than 0.6 in the scene semantic feature map are screened, and the three-dimensional spatial coordinates are calculated based on the depth values ​​at the corresponding positions; the semantic attributes of the feature points are associated with the three-dimensional coordinates to generate a three-dimensional feature point cloud with spatial position and semantic information.

[0045] S3.2: Preprocess the feature point cloud.

[0046] Specifically, statistical filtering is used to remove outliers, and the voxel downsampling method is used to reduce the point cloud density, retaining the point at the center of gravity position in each voxel; the normal vector is estimated for the downsampled point cloud and the direction of the normal vector is determined; at the same time, the point cloud is corrected for posture to eliminate the tilt caused by camera installation errors.

[0047] S3.3: Perform spatial clustering and segmentation on the preprocessed feature point cloud to establish local feature areas.

[0048] Specifically, a region growing algorithm based on normal vector consistency constraint is adopted: first, the point with the smallest curvature is selected as the seed point, and a spherical search area is constructed with the seed point as the center. Points in the search area with a normal vector angle less than 30 degrees and a distance less than 0.1 meters are merged into the current area; the above process is repeated until all points are assigned region labels; finally, the regions are merged based on the cosine similarity of the average semantic features between regions, and the similarity threshold is set to 0.8 to obtain the local feature region.

[0049] S3.4: Construct an octree spatial index based on the local feature area, divide the three-dimensional grid units, and adaptively adjust the grid resolution and octree depth according to the complexity of the scene.

[0050] Specifically, the size of the bounding box of the root node of the initial octree is determined according to the range of the input point cloud; the space is recursively divided from top to bottom, and the division continues when the number of points in the node is greater than the number threshold N or the current division depth is less than the preset maximum depth D. In this embodiment, N=100 and D=8; during the division process, the point cloud density and average curvature of the node are calculated in real time. If the point cloud density of the node exceeds the preset density threshold or the average curvature exceeds the preset curvature threshold, the node is marked as a complex area; for nodes in complex areas, the octree division depth is increased by 2 additional layers on the basis of the original maximum depth, and finally a spatial grid division with adaptive resolution is obtained.

[0051] S3.5: In each 3D mesh cell, a Poisson reconstruction algorithm is used to reconstruct the 3D surface mesh.

[0052] Specifically, the Poisson equation is constructed based on the position and normal vector information of the feature points, the Poisson equation is solved, and the marching cubes algorithm is used to extract the isosurface, and the position with the isovalue of 0 is the reconstructed surface; the mesh of the reconstructed surface is simplified, and the number of triangles is controlled within 50% of the original number.

[0053] S3.6: splice the three-dimensional surface meshes of adjacent three-dimensional mesh units along the boundaries to construct a three-dimensional scene mesh model.

[0054] In the specific implementation, firstly, the boundary point set is extracted in the overlapping area of ​​adjacent grid cells, and the width of the overlapping area is set to 10% of the grid cell size; the boundary points are aligned using the iterative closest point algorithm to obtain the boundary correspondence; based on the boundary correspondence, the grid is stitched, the duplicate vertices are deleted and the local triangle connection relationship is reconstructed; finally, the grid in the boundary area is Laplace smoothed, the smoothing weight coefficient is set to 0.3, and the number of iterations is controlled to 3 times to ensure the smooth transition of the grid.

[0055] S4: Extract the spatial topological features of each grid node in the three-dimensional scene grid model and calculate the association weights between the grid nodes.

[0056] Specifically, the following steps are included:

[0057] S4.1: Construct a K-nearest neighbor graph for the grid nodes in the three-dimensional scene grid model and record the local connection relationship of each grid node.

[0058] Specifically, for each grid node, K nearest neighbor nodes are searched in the Euclidean space. The K value is set adaptively according to the grid resolution, and is generally 6 to 12. An adjacency matrix is ​​constructed to store the connection relationship between nodes, and the corresponding edge weights are recorded as the Euclidean distance between nodes.

[0059] S4.2: Calculate the geometric descriptor of each mesh node, the geometric descriptor includes the normal vector, principal curvature and local surface shape coefficient of the mesh node.

[0060] Preferably, principal component analysis is performed on triangular facets in the local neighborhood of the node to calculate normal vectors and principal curvatures; shape index and curvature change rate are calculated based on the principal curvature to obtain local surface shape coefficients; and these features are combined to form a geometric descriptor of the node.

[0061] S4.3: Based on the local connection relationship, a graph convolutional neural network is used to aggregate the features of the grid nodes to generate a high-order topological feature representation of the grid nodes.

[0062] Specifically, a three-layer graph convolutional network is constructed, and each layer includes two steps: feature propagation and feature update. An adaptive weight mechanism based on the local curvature and normal vector of the node is introduced. In the feature propagation stage, message passing is based on the improved adjacency matrix, and layer normalization and gated update units are added. In the feature update stage, node features are fused through nonlinear transformation. A multi-scale residual connection strategy is designed to retain the underlying geometric features. A residual connection mechanism is used to prevent over-smoothing, and a dynamic feature dimension adjustment mechanism is introduced to finally output an adaptive node feature vector.

[0063] S4.4: Establish a feature similarity matrix of grid nodes and calculate the spatial correlation based on the Euclidean distance between nodes.

[0064] Preferably, cosine similarity is used to calculate the similarity between node feature vectors; the feature similarity and the normalized spatial distance are combined in a weighted manner to obtain a comprehensive spatial association measure, wherein the weight coefficient is determined by cross-validation.

[0065] S4.5: Based on the high-order topological feature representation and spatial correlation, the attention mechanism is used to calculate the association weights between grid nodes.

[0066] Specifically, high-order topological features are mapped to query and key-value spaces, attention scores are calculated, and the expressiveness of the attention mechanism is enhanced by introducing node geometric constraints and topological consistency terms. The attention score is multiplied element-by-element by the spatial correlation, and the final correlation weight is obtained by normalization through the Softmax function. This weight reflects the strength of geometric and topological correlation between nodes.

[0067] Among them, the design of the attention mechanism considers the following key factors: feature space mapping, geometric constraints, topological consistency, and multi-scale feature fusion.

[0068] In addition, it should be noted that the spatial topological features include the local connection relationship of grid nodes, geometric descriptors, high-order topological feature representations, and spatial correlation.

[0069] S5: Construct a scene dynamic knowledge graph based on spatial topological features and association weights to model and analyze abnormal events.

[0070] Specifically, the following steps are included:

[0071] S5.1: Define node attribute vectors based on spatial topological features, and construct edge attribute vectors based on association weights to form the basic elements of the scene knowledge graph.

[0072] Specifically, the spatial topological features and semantic features are concatenated to form a node attribute vector; the association weight is combined with the relative spatial position and motion features between node pairs to form an edge attribute vector; each node and edge also contains a timestamp attribute to record the status update time.

[0073] S5.2: Use a graph database to store the basic elements of the scene knowledge graph and establish a time-series index mapping table.

[0074] Preferably, the attribute vectors of nodes and edges are stored in the form of key-value pairs, using Neo4j storage, and the corresponding attribute vectors of nodes and edges are stored with unique identifiers as keys; a time series index table is constructed to record the evolution of the graph structure over time; a sliding time window mechanism is set to retain the historical status information of the most recent N frames, and the N value is determined according to the system memory capacity.

[0075] S5.3: Use the attention graph neural network to dynamically evolve the scene knowledge graph and generate an embedded representation of the scene state.

[0076] Specifically, a graph neural network containing temporal attention and spatial attention is constructed, including a temporal attention module and a spatial attention module. The specific implementation method of the temporal attention module is as follows: first, the historical state sequence of the node is encoded, and the states of different time steps are nonlinearly weighted fused through learnable attention weights, and residual connections and layer normalization techniques are combined. The specific implementation method of the spatial attention module is as follows: the attention coefficient between nodes is calculated based on the edge attribute vector, the interaction weights between nodes are dynamically adjusted through attention mechanism learning, and the node states are iteratively updated using the message propagation paradigm. Through the above methods, a multi-scale state embedding representation that reflects the dynamic changes of the scene is finally obtained.

[0077] S5.4: Design a multi-task learning module to input the scene state embedding representation into the anomaly classifier and the localization regressor respectively.

[0078] Preferably, the network structure design of the anomaly classifier includes: the input layer receives the scene state embedding vector, passes through three layers of fully connected neural network, adds batch normalization and dropout regularization to each layer, and finally outputs the probability distribution of the anomaly category through the softmax activation function; the positioning regressor adopts a twin network structure, including two parallel branch networks, which respectively predict the spatial range and occurrence time of the abnormal target. Each branch network consists of a convolution layer, a pooling layer and a fully connected layer, and the feature extraction capability is enhanced through residual connection and attention mechanism. The two subtasks realize feature reuse by sharing the encoding layer parameters.

[0079] S5.5: Fusion the outputs of the anomaly classifier and the localization regressor to generate a spatiotemporal feature description of the anomaly event.

[0080] Specifically, the classification results and positioning results are confidence-weighted and fused, and a confidence threshold is set to filter the detection results; for the detected abnormal events, their key attributes such as category, location, impact range, duration, etc. are recorded; an abnormal event database is established to support subsequent statistical analysis and retrieval queries.

[0081] Furthermore, the present embodiment also provides an environmental modeling system for realizing intelligent security, including a data acquisition module, which is used to use multi-eye vision sensors and depth sensors to collect visual images and depth data of the monitoring area, and generate a monitoring scene feature matrix; a feature extraction module, which is used to use a deep convolutional neural network to perform hierarchical feature extraction on the monitoring scene feature matrix, and introduce a spatial-channel dual attention module to calibrate important area and channel features, and construct a scene semantic feature map; a three-dimensional modeling module, which is used to project the scene semantic feature map into a three-dimensional monitoring space according to the depth data, and construct a three-dimensional scene grid model; a topological feature extraction module, which is used to extract the spatial topological features of each grid node in the three-dimensional scene grid model, and calculate the association weights between the grid nodes; an abnormal module, which is used to construct a scene dynamic knowledge graph based on the spatial topological features and the association weights, and model and analyze abnormal events.

[0082] In summary, the present invention realizes the precise generation of monitoring scene feature matrix through the collaborative acquisition of multi-eye vision sensors and depth sensors, combined with information entropy weighted normalization and principal component analysis, effectively solving the problems of information redundancy and feature coupling in traditional visual data processing; introduces the space-channel dual attention module, and realizes the adaptive extraction and enhancement of scene semantic features by constructing a multi-scale convolutional hierarchical network and residual connection units, which significantly improves the robustness and discrimination of feature expression; constructs a three-dimensional scene grid model based on depth data, extracts grid node topological features, and constructs a dynamic knowledge graph based on attention graph neural network, innovatively extending scene understanding from a two-dimensional plane to a three-dimensional space, providing a more comprehensive and in-depth technical path for the precise modeling and intelligent analysis of abnormal events.

[0083] Example 2, reference Figure 1-2 , which is the second embodiment of the present invention, and this embodiment provides an environmental modeling method for realizing intelligent security. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0084] To verify the effectiveness of the proposed intelligent security environment modeling method, the research team conducted comprehensive simulation experiments in laboratory environments and real monitoring scenarios. The experimental platform consists of a multi-view visual sensor array and a depth sensor, with a total monitoring area of ​​300 square meters, covering typical scenarios such as office areas, corridors, and entrances and exits. The sensor configuration includes 6 RGB cameras with a 120° field of view and 3 depth sensors, with a sampling frequency of 60Hz and a spatial resolution of 2048×1536 pixels. The data collection process lasted 30 days, and a total of 1.2TB of multimodal sensor data was collected, including different lighting conditions, personnel density, and activity scenes.

[0085] In the feature extraction stage, by analyzing the 32×32 pixel image blocks, the research team found that the feature robustness of this method is significantly better than the traditional image block division method. The average accuracy of image feature extraction reached 92.5%, which was 15.3% and 10.7% higher than the traditional HOG and SIFT feature extraction methods, respectively. The adaptive threshold strategy for deep data region segmentation increased the segmentation accuracy to 96.1%, effectively reducing background noise interference.

[0086] The feature extraction experiment of multi-scale convolutional neural network shows that the introduction of spatial-channel dual attention module significantly enhances the scene semantic understanding ability. On the standard public dataset, the model's target detection average precision (mAP) reaches 87.6%, which is 12.4 percentage points higher than the baseline convolutional neural network. The spatial attention module has an accuracy of 91.3% in locating key areas, and the channel attention module can effectively highlight semantically related feature channels.

[0087] The 3D scene reconstruction experiment selected five surveillance scenes of different complexity, including an open office area, a narrow corridor, and a multi-branch aisle. The average accuracy of point cloud reconstruction reached 93.7%, and the octree adaptive partitioning strategy increased the mesh resolution of complex areas by 45.2%. The Poisson reconstruction algorithm significantly improved the surface continuity, the number of triangles was controlled within 50% of the original point cloud, and the geometric detail fidelity of the reconstructed model reached 89.5%.

[0088] As shown in Table 1, the comparative analysis between the present invention and the existing intelligent security environment modeling technology reveals the advantages of the method:

[0089] Table 1 Performance comparison between the method of the present invention and the traditional method

[0090] Comparison Dimensions Method of the present invention Traditional methods Feature extraction accuracy 92.5％ 76.8％ Scene reconstruction accuracy 93.7％ 78.4％ Abnormal event detection recall rate 89.4％ 72.6％ Computational complexity Low high Real-time performance high middle

[0091] The knowledge graph construction and abnormal event analysis experiments were conducted in five different scenarios, including crowd gathering, suspicious behavior, and unauthorized access. The anomaly classification accuracy of the multi-task learning module reached 88.6%, and the spatial error of the positioning regressor was controlled within 0.15 meters. Through the attention graph neural network, the system can capture dynamic scene changes in real time, and the average response time of abnormal event detection does not exceed 200 milliseconds.

[0092] Experimental results show that the intelligent security environment modeling method has significant technical advantages in multimodal feature fusion, scene understanding and anomaly detection, and provides an efficient and accurate environment modeling solution for intelligent security systems.

[0093] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An environmental modeling method for realizing intelligent security, characterized in that: include, Collect visual images and depth data of the monitoring area to generate a monitoring scene feature matrix; A deep convolutional neural network is used to extract hierarchical features from the monitoring scene feature matrix, and a spatial-channel dual attention module is introduced to calibrate important area and channel features to construct a scene semantic feature map; Projecting the scene semantic feature map into a three-dimensional monitoring space according to the depth data to construct a three-dimensional scene mesh model; Extracting spatial topological features of each grid node in the three-dimensional scene grid model, and calculating association weights between grid nodes; A scene dynamic knowledge graph is constructed based on the spatial topological features and associated weights to model and analyze abnormal events.

2. The method for implementing intelligent security environment modeling according to claim 1, characterized in that: Generating the monitoring scene feature matrix comprises the following steps: Perform image block segmentation on the visual image and extract the color distribution histogram and texture feature vector of each image block; Performing region segmentation on the depth data, and calculating a depth gradient histogram and a curvature feature vector for each region; Performing information entropy weighted normalization preprocessing on the color distribution histogram, the texture feature vector, the depth gradient histogram, and the curvature feature vector to obtain a standardized feature fusion matrix; Performing dimensionality reduction and feature decoupling on the standardized feature fusion matrix through principal component analysis to extract main information components; The main information components are reconstructed into a monitoring scene feature matrix.

3. The method for implementing intelligent security environment modeling according to claim 1, characterized in that: The construction of the scene semantic feature map comprises the following steps: Constructing a multi-scale convolutional hierarchical network to perform feature decomposition on the monitoring scene feature matrix through convolution kernels of different scales; Establishing a spatial attention module, calculating the spatial position correlation score after the feature decomposition, and generating a spatial weight matrix; Establishing a channel attention module, calculating the channel correlation score after the feature decomposition, and generating a channel weight matrix; Mapping the spatial weight matrix and the channel weight matrix to the feature space respectively, and obtaining a dual attention weight map through nonlinear transformation; The feature decomposition result is dynamically enhanced according to the dual attention weight map to generate a scene semantic feature map.

4. The method for implementing intelligent security environment modeling according to claim 1, characterized in that: The construction of the three-dimensional scene mesh model comprises the following steps: Perform three-dimensional coordinate back-projection on the feature points in the scene semantic feature map using the depth data to generate a feature point cloud; Preprocessing the feature point cloud; Perform spatial clustering and segmentation on the preprocessed feature point cloud to establish local feature areas; Building an octree spatial index based on the local feature area, dividing the three-dimensional grid unit, and adaptively adjusting the grid resolution and octree depth according to the complexity of the scene; In each 3D mesh unit, the Poisson reconstruction algorithm is used to reconstruct the 3D surface mesh; The three-dimensional surface meshes of adjacent three-dimensional mesh units are spliced ​​along the boundaries to construct a three-dimensional scene mesh model.

5. The method for implementing intelligent security environment modeling according to claim 1, characterized in that: The calculation of the association weights between grid nodes comprises the following steps: Construct a K-nearest neighbor graph for the grid nodes in the 3D scene grid model and record the local connection relationship of each grid node; Calculating a geometric descriptor for each mesh node, wherein the geometric descriptor includes a normal vector, a principal curvature, and a local surface shape coefficient of the mesh node; Based on the local connection relationship, a graph convolutional neural network is used to aggregate the features of the grid nodes to generate a high-order topological feature representation of the grid nodes; Establish the feature similarity matrix of grid nodes and calculate the spatial correlation by combining the Euclidean distance between nodes; According to the high-order topological feature representation and the spatial correlation, an attention mechanism is used to calculate the correlation weights between grid nodes; The spatial topological features include local connection relationships of grid nodes, geometric descriptors, high-order topological feature representations, and spatial correlation.

6. The environmental modeling method for realizing intelligent security according to claim 1, characterized in that: The construction of the scene dynamic knowledge graph includes the following steps: Define node attribute vectors based on spatial topological features, and construct edge attribute vectors based on association weights to form the basic elements of the scene knowledge graph; Using a graph database to store basic elements of the scene knowledge graph and establishing a time series index mapping table; Using an attention graph neural network to dynamically evolve the scene knowledge graph to generate an embedded representation of the scene state; Design a multi-task learning module to input the scene state embedding representation into the anomaly classifier and the positioning regressor respectively; The output results of the anomaly classifier and the positioning regressor are integrated to generate a spatiotemporal feature description of the abnormal event.

7. The environmental modeling method for realizing intelligent security according to claim 3, characterized in that: The feature decomposition of the monitoring scene feature matrix by using convolution kernels of different scales comprises the following steps: Constructing a multi-scale convolutional hierarchical network, wherein the multi-scale convolutional hierarchical network includes parallel convolution branches, and the convolution kernel sizes of the parallel convolution branches are configured according to an increasing rule; Inputting the monitoring scene feature matrix into the parallel convolution branch, and obtaining convolution features through convolution kernels of different scales; A residual connection unit is provided in the parallel convolution branch, wherein the residual connection unit combines the input feature with the convolution feature; Perform channel splicing on the features output by the parallel convolution branches to obtain multi-scale fusion features; The multi-scale fusion features are processed using a singular value decomposition algorithm to generate a feature decomposition result.

8. An environment modeling system for realizing intelligent security, based on the environment modeling method for realizing intelligent security according to any one of claims 1 to 7, characterized in that: Also includes, The data acquisition module is used to collect visual images and depth data of the monitoring area using multi-eye visual sensors and depth sensors to generate a monitoring scene feature matrix; A feature extraction module is used to perform hierarchical feature extraction on the monitoring scene feature matrix using a deep convolutional neural network, and introduce a space-channel dual attention module to calibrate important area and channel features to construct a scene semantic feature map; A three-dimensional modeling module, used to project the scene semantic feature map into a three-dimensional monitoring space according to the depth data to construct a three-dimensional scene mesh model; A topological feature extraction module is used to extract the spatial topological features of each grid node in the three-dimensional scene grid model and calculate the association weights between the grid nodes; The anomaly module is used to construct a scene dynamic knowledge graph based on the spatial topological features and associated weights, and to model and analyze abnormal events.

Images