A method and system for intelligent three-dimensional security monitoring in smart parks

Abstract

This invention relates to an intelligent three-dimensional security monitoring method and system for smart parks, belonging to the field of security monitoring technology. The method includes constructing a multi-dimensional three-dimensional perception network to acquire and preprocess multimodal heterogeneous data; post-processing the data in parallel using a spatiotemporal graph neural network and a Bayesian network, and adaptively fusing them at the decision level through a gating fusion mechanism to construct a security monitoring model; inputting the preprocessed data into the security monitoring model to generate dynamic feature vectors and rule confidence vectors; and weighted fusing the two vectors through an adaptive gating fusion mechanism at the decision level to output real-time alarm levels, event type probability distributions, and evidence summaries.

Description

Technical Field

[0001] This invention relates to the field of security monitoring technology, and in particular to an intelligent three-dimensional security monitoring method and system for smart parks. Background Technology

[0002] Existing park security systems generally face multi-dimensional technical bottlenecks, which are severely restricting the effectiveness of smart park security operations. Recent industry reports on security monitoring show that over half of security incidents in my country's industrial parks are directly related to delayed security system responses, while nearly half of property losses in science and technology parks are closely related to the waste of resources caused by false alarms. These figures not only reveal the failure of security systems at critical moments but also highlight deep-seated structural flaws in the existing technological system, making it difficult to meet the increasingly complex security needs of parks.

[0003] A thorough analysis of existing security systems reveals that their simplistic technical architecture is the key factor leading to the aforementioned problems. Statistics show that a staggering 89% of existing park security systems rely excessively on planar video surveillance. This reliance has resulted in numerous drawbacks. Traditional cameras typically have an effective monitoring height limited to 3 to 10 meters, leaving the system lacking the necessary detection capabilities for high-altitude risks such as drone intrusions and leaks in overhead pipelines, creating a "vertical dimension gap." Simultaneously, only a very small number of parks have deployed underground sensing equipment, resulting in the inability to monitor underground hazards such as cable trench theft and hazardous chemical pipeline leaks in real time, creating an "underground dimension gap." Even more worrying is that the coverage rate of existing security systems in concealed areas such as building mezzanines and equipment rooms is often less than half, making these areas breeding grounds for illegal intrusions and creating "blind spots in concealed spaces." This coexistence of vertical dimension gaps, underground dimension gaps, and blind spots in concealed spaces constitutes a situation of "triple protection gaps."

[0004] Single-layer security monitoring, whether at high altitude, on the ground, or underground, cannot provide comprehensive security. This planar and fragmented monitoring model makes the entire security system appear to have multiple weak points; the failure of any one link could lead to the collapse of the entire security system. Therefore, existing security systems not only have obvious security vulnerabilities but also struggle to meet the urgent needs of smart parks for security, efficiency, and intelligence, necessitating the introduction of more comprehensive and multi-dimensional advanced technological solutions. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, the main objective of this invention is to provide an intelligent three-dimensional security monitoring method and system for smart parks. By constructing a multi-dimensional perception network, innovatively integrating spatiotemporal graph neural networks and Bayesian networks, and designing an adaptive decision-making mechanism, efficient, accurate and interpretable park security monitoring is achieved.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent three-dimensional security monitoring method for smart parks, comprising the following steps:

[0007] A multi-dimensional three-dimensional sensing network is constructed using UAV sensing network, ground fixed sensing network and mobile IoT sensing network. Based on the three-dimensional sensing network, multimodal heterogeneous data is acquired and preprocessed.

[0008] A security monitoring model is constructed by parallelizing spatiotemporal graph neural networks and Bayesian networks and adaptively fusing them at the decision layer.

[0009] The preprocessed multimodal heterogeneous data is input into the security monitoring model. The dynamic time series data in the preprocessed multimodal heterogeneous data is processed by the spatiotemporal graph neural network to obtain dynamic feature vectors. The state event data is processed by the differentiable conditional probability model. Rule reasoning is performed based on the conditional probability table generated by the expert database rules, and the rule confidence vector is output.

[0010] The design employs an adaptive gating fusion mechanism at the decision-making level to adaptively weight and fuse the dynamic feature vector and rule confidence vector, generating multi-level decision outputs, including real-time alarm levels, event type probability distributions, and evidence summaries, which serve as the security monitoring results.

[0011] The construction of the multi-dimensional stereo perception network includes:

[0012] The drone sensing network adopts a heterogeneous cluster collaborative architecture, including visible light drones, infrared temperature measurement drones, and laser scanning drones;

[0013] The ground-based fixed sensor network deploys multispectral video monitoring nodes;

[0014] The mobile IoT sensing network tracks personnel using UWB / BLE dual-mode positioning tags.

[0015] The process of processing the dynamic temporal data in the preprocessed multimodal heterogeneous data using a spatiotemporal graph neural network to obtain dynamic feature vectors includes the following steps:

[0016] The key points in the monitored area are used as nodes, the spatiotemporal relationships between nodes are used as edges, and the adjacency matrix obtained from the dynamic time series data is used to construct a spatiotemporal graph.

[0017] The spatiotemporal graph convolutional layer in the spatiotemporal graph neural network extracts the spatial features of the spatiotemporal graph, and the spatiotemporal features are obtained through the gated temporal convolutional layer; the encoder of the spatiotemporal graph neural network encodes the spatiotemporal features to generate dynamic feature vectors.

[0018] The process of processing state event data using a differentiable conditional probability model, performing rule inference based on a conditional probability table generated from expert database rules, and outputting a rule confidence vector includes the following steps:

[0019] By treating security-related events as nodes in a Bayesian network and using probabilistic dependencies defined by expert database rules as edges, a Bayesian network topology is constructed.

[0020] Based on the network topology and initial conditional probability table defined by the expert database rules, a differentiable conditional probability model is constructed, in which the conditional probability table of each non-root node is obtained through neural network parameterization.

[0021] The event and state record data in the state event data are mapped to the corresponding leaf nodes in the Bayesian network topology, and the observation data of the events and states are encoded into the embedding vectors of the corresponding leaf nodes; the corresponding leaf nodes and the embedding vectors of the corresponding leaf nodes are used as evidence vectors.

[0022] Using the evidence vector and the differentiable conditional probability model, probabilistic reasoning is performed to obtain the posterior probability distribution of the target event node;

[0023] The confidence level of the target event is extracted based on the posterior probability distribution, and a rule confidence vector is generated.

[0024] Each non-root node is modeled with a conditional probability table using a hierarchical neural network structure, including:

[0025] The first layer of the sub-network groups the parent node states according to logical correlation and outputs the intermediate representation of each group of parent nodes;

[0026] The second-layer fusion network takes the intermediate representations as input and outputs the state probability distribution of the node.

[0027] The probabilistic reasoning employs a differentiable approximate reasoning algorithm.

[0028] The method for generating the rule confidence vector is as follows:

[0029] The target event node is a binary state, and its positive class state probability is taken as the confidence level.

[0030] The target event node is a multi-state entity, and the confidence level of its complete state probability vector is taken.

[0031] The gating fusion mechanism includes:

[0032] The dynamic feature vector and the rule confidence vector are dimensionally aligned and concatenated to generate a fused input vector.

[0033] An adaptive weight is generated for a dynamic feature vector and a rule confidence vector using a gated neural network. The gated neural network consists of two fully connected layers: the first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the Softmax function. The adaptive weight consists of a dynamic feature vector weight α and a rule confidence vector weight β, which satisfy α + β = 1.

[0034] Based on the adaptive weights, the dynamic feature vector and the rule confidence vector are weighted and fused to generate a fused feature vector;

[0035] A multi-task decision head is introduced, and the fused feature vector is input into the multi-task decision head to generate multi-level decision outputs.

[0036] The multi-task decision head includes:

[0037] The continuously generated alarm level values ​​are output through a fully connected layer and used as the alarm level generation branch.

[0038] The event type probability distribution is output through a fully connected layer and a Softmax function, serving as the event classification branch.

[0039] Key evidence is extracted from the dynamic feature vector and rule confidence vector based on attention weights to generate a structured text summary, which serves as the evidence summary generation branch.

[0040] The evidence digest generation includes:

[0041] Obtain the attention weights of the fused feature vector, the dynamic feature vector, and the rule confidence vector;

[0042] The top N features with the highest attention weights are selected as key evidence.

[0043] A structured text summary is generated by filling key evidence into a predefined template; wherein the predefined template includes an event type field, a confidence level field, and an evidence source field.

[0044] Compared with existing technologies, the beneficial effects of this invention are as follows: By integrating multi-dimensional sensing networks such as drones, fixed ground devices, and mobile IoT, this invention solves the "triple protection gap" problem in traditional security systems, achieving comprehensive coverage and real-time monitoring of high-altitude, ground-based, underground, and concealed spaces within the park. This significantly enhances the breadth and depth of security protection and effectively prevents various security incidents. By employing a spatiotemporal graph neural network to process dynamic time-series data, it accurately captures the evolution patterns and correlations of events in time and space, extracting high-dimensional dynamic features. Simultaneously, by combining a differentiable conditional probability model with a Bayesian network for rule-based reasoning, it effectively integrates expert experience and knowledge into the model, achieving parallel processing and deep fusion of data-driven and rule-driven approaches. This improves the understanding of complex scenarios and the accuracy of decision-making. Furthermore, the designed decision-level adaptive gating fusion mechanism can dynamically adjust the weights of data-driven features and rule confidence based on the different characteristics of the input data, generating more accurate and reliable multi-level decision outputs. This ensures optimal responses under different security situations, improving the intelligence and adaptability of decision-making.

[0045] To enhance the interpretability and credibility of security monitoring, this invention utilizes the rule-based reasoning process of Bayesian networks and the dynamic adjustment of weights in the gating fusion mechanism. This ensures that the final decision is based not only on data patterns but also incorporates expert rules, making the decision-making process more transparent. In particular, the evidence summary generation module can extract key evidence from massive amounts of data and present it in a structured manner, providing security personnel with intuitive and credible decision-making support, significantly improving the interpretability and user trust of the security system. This invention not only outputs traditional alarm signals but also generates continuous alarm levels, fine-grained event type probability distributions, and structured evidence summaries. This refined output method enables security managers to grasp the event situation more comprehensively and accurately, providing more specific and powerful support for subsequent resource allocation and emergency response, thereby improving overall security operational efficiency.

[0046] In summary, this invention, through technological innovation, effectively solves many pain points in existing smart park security monitoring, achieving a leap from passive response to proactive defense and from single monitoring to intelligent decision-making, providing strong technical support for the safe and stable operation of smart parks. Attached Figure Description

[0047] The accompanying drawings, which are provided to further illustrate this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application.

[0048] Figure 1 This is a schematic diagram of the process of this invention.

[0049] Figure 2This is a schematic diagram of the architecture of this invention.

[0050] Figure 3 This is a schematic diagram of the three-dimensional sensing network of the present invention.

[0051] Figure 4 This is a schematic diagram of the spatiotemporal graph neural network processing flow of the present invention.

[0052] Figure 5 This is a schematic diagram of the Bayesian network processing flow of the present invention.

[0053] Figure 6 This is a schematic diagram of the hierarchical conditional probability table structure of the present invention.

[0054] Figure 7 This is a schematic diagram of the gating fusion process of the present invention.

[0055] Figure 8 This is a schematic diagram of the multi-task decision-making process of the present invention. Detailed Implementation

[0056] Smart parks integrate numerous advanced facilities and complex activities, posing unprecedented challenges to security monitoring. Traditional security systems often rely on single monitoring methods such as fixed cameras, resulting in blind spots, data silos, and delayed analysis. Single-layer security monitoring, whether at high altitudes, on the ground, or underground, struggles to provide comprehensive security. This planar and fragmented monitoring model leaves the entire security system vulnerable, with the failure of any one link potentially causing the entire system to collapse. Therefore, existing security systems not only present significant security risks but also fail to meet the urgent needs of smart parks for security, efficiency, and intelligence, necessitating the introduction of more comprehensive and multi-dimensional advanced technological solutions.

[0057] To address these challenges, this paper proposes an intelligent three-dimensional security monitoring method for smart parks. This method breaks down data barriers by constructing a multi-dimensional three-dimensional perception network, achieving full-domain coverage of multi-dimensional monitoring. By deeply fusing spatiotemporal graph neural networks and Bayesian networks, it leverages the data-driven deep learning capabilities of monitoring data to capture complex patterns, while combining rule-driven expert knowledge to improve decision reliability. Through adaptive gating fusion and multi-task decision head design, it generates rich, reliable, and easily understandable security monitoring results. (See also...) Figures 1 to 8 It can be seen that the specific steps include:

[0058] A multi-dimensional, three-dimensional sensing network is constructed by utilizing drone sensing networks, ground-based fixed sensing networks, and mobile IoT sensing networks. This process includes the following steps:

[0059] The drone perception network adopts a heterogeneous cluster collaborative architecture, including visible light drones, infrared temperature measurement drones, and laser scanning drones;

[0060] Deploy multispectral video surveillance nodes using a ground-based fixed sensor network;

[0061] Mobile IoT sensing networks track people using dual-mode positioning tags with both UWB and BLE.

[0062] Based on the constructed multi-dimensional stereo perception network, multimodal heterogeneous data is acquired and preprocessed to obtain preprocessed multimodal heterogeneous data;

[0063] A security monitoring model is constructed by parallelizing spatiotemporal graph neural networks and Bayesian networks and adaptively fusing them at the decision layer.

[0064] The preprocessed multimodal heterogeneous data is input into the security monitoring model. The dynamic time-series data within the preprocessed multimodal heterogeneous data is then processed using a spatiotemporal graph neural network to obtain dynamic feature vectors, specifically:

[0065] The key points in the monitored area are used as nodes, the spatiotemporal relationships between nodes are used as edges, and the adjacency matrix obtained from the dynamic time series data is used to construct a spatiotemporal graph.

[0066] The spatiotemporal graph convolutional layer in the spatiotemporal graph neural network extracts the spatial features of the spatiotemporal graph, and the spatiotemporal features are obtained through the gated temporal convolutional layer; the encoder of the spatiotemporal graph neural network encodes the spatiotemporal features to generate dynamic feature vectors.

[0067] Then, the state event data is processed through a differentiable conditional probability model, and rule reasoning is performed based on the conditional probability table generated by the expert database rules to output the rule confidence vector; specifically, security-related events are used as nodes of the Bayesian network, and the probability dependencies defined by the expert database rules are used as edges to construct the Bayesian network topology.

[0068] Based on the network topology and initial conditional probability table defined by the expert database rules, a differentiable conditional probability model is constructed, in which the conditional probability table of each non-root node is obtained through neural network parameterization; for each non-root node, a hierarchical neural network structure is used to model its conditional probability table, including the following steps:

[0069] The first layer of the sub-network groups the parent node states according to logical correlation and outputs the intermediate representation of each group of parent nodes;

[0070] The second-layer fusion network takes the intermediate representations as input and outputs the state probability distribution of the node.

[0071] Probabilistic reasoning employs a differentiable approximate reasoning algorithm.

[0072] The event and state record data in the state event data are mapped to the corresponding leaf nodes in the Bayesian network topology, and the observation data of the events and states are encoded into the embedding vectors of the corresponding leaf nodes; the corresponding leaf nodes and the embedding vectors of the corresponding leaf nodes are used as evidence vectors.

[0073] Using the evidence vector and the differentiable conditional probability model, probabilistic reasoning is performed to obtain the posterior probability distribution of the target event node;

[0074] The confidence level of the target event is extracted based on the posterior probability distribution, and a regular confidence vector is generated. The regular confidence vector is generated as follows:

[0075] The target event node is a binary state, and its positive class state probability is taken as the confidence level.

[0076] The target event node is a multi-state entity, and the confidence level of its complete state probability vector is taken.

[0077] In the final decision-making stage, an adaptive gating fusion mechanism needs to be designed to adaptively weight and fuse dynamic feature vectors and rule confidence vectors, generating multi-level decision outputs, including real-time alarm levels, event type probability distributions, and evidence summaries, as the security monitoring results. The design of the gating fusion mechanism includes:

[0078] The dynamic feature vector and the rule confidence vector are dimensionally aligned and concatenated to generate a fused input vector.

[0079] An adaptive weight is generated by a gated neural network to form a dynamic feature vector and a regular confidence vector. The gated neural network consists of two fully connected layers. The first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the Softmax function. The adaptive weight consists of a dynamic feature vector weight α and a regular confidence vector weight β, which satisfy α + β = 1.

[0080] Based on adaptive weights, the dynamic feature vector and the rule confidence vector are weighted and fused to generate a fused feature vector;

[0081] A multi-task decision head is introduced, and the fused feature vector is input into the multi-task decision head to generate multi-level decision outputs.

[0082] The introduced multi-task decision head specifically includes:

[0083] The continuously generated alarm level values ​​are output through a fully connected layer and used as the alarm level generation branch.

[0084] The event type probability distribution is output through a fully connected layer and a Softmax function, serving as the event classification branch.

[0085] Based on attention weights, key evidence is extracted from the dynamic feature vector and rule confidence vector to generate a structured text summary, which serves as the evidence summary generation branch.

[0086] The evidence summary generation includes:

[0087] Obtain the attention weights of the fused feature vector, dynamic feature vector, and rule confidence vector;

[0088] The top N features with the highest attention weights are selected as key evidence.

[0089] The key evidence is populated to generate a structured text summary based on a predefined template; the predefined template includes an event type field, a confidence level field, and an evidence source field.

[0090] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0091] Example 1: The following is an example of an intelligent three-dimensional security monitoring method for smart parks, incorporating detailed technical aspects to better understand the implementation methods and application scenarios of each stage. In this smart park security monitoring system, comprehensive monitoring of the park environment is first achieved through a multi-dimensional three-dimensional sensing network. (See also...) Figures 1 to 8 It can be seen that the perceptual network consists of the following parts:

[0092] The first step is to construct a three-dimensional perception network, which specifically includes:

[0093] The drone sensing network is a heterogeneous swarm sensing network constructed through the collaborative work of multiple types of drones. Visible light drones, infrared thermography drones, and laser scanning drones are each responsible for different data acquisition tasks.

[0094] Visible light drones are responsible for acquiring real-time image data and conducting real-time monitoring.

[0095] Infrared temperature measurement drones are responsible for monitoring temperature changes, especially in areas of the park where temperature variations are significant.

[0096] Laser scanning drones are used for high-precision mapping to generate 3D models of the park, providing detailed data support for spatial layout and security protection.

[0097] A ground-based fixed sensor network is deployed, with multispectral video surveillance nodes to cover the main areas within the park. These sensors perform intelligent analysis by fusing video and infrared data.

[0098] The mobile IoT sensing network uses UWB / BLE dual-mode positioning tags for personnel tracking, which can accurately identify the movement trajectory of people within the park.

[0099] The collected multimodal data, including images, infrared video, and temperature data, undergoes preprocessing. Noise removal and data standardization techniques are used to ensure the quality and accuracy of the data input into subsequent models.

[0100] The fusion of spatiotemporal graph neural networks and Bayesian networks begins with the construction and dynamic feature extraction of the spatiotemporal graph neural network. Key locations within the park, such as gates and monitoring points, are used as nodes to construct a spatiotemporal graph. The connections between nodes reflect their temporal and spatial correlations. Spatial features of the nodes are extracted using spatiotemporal graph convolutional layers, while temporal features between nodes are extracted using gated temporal convolutional layers. An encoder encodes these spatiotemporal features to generate dynamic feature vectors. These dynamic feature vectors contain key information from the time-series data, reflecting real-time dynamic changes within the park.

[0101] Security incidents such as fires and intrusions are treated as nodes in a Bayesian network. The dependencies between nodes are determined by rules from an expert database, which constructs the topology of the Bayesian network.

[0102] Based on Bayesian networks, the conditional probability table of each non-root node is parameterized using a neural network. The posterior probability of the event is calculated through inference, and the rule confidence vector is output.

[0103] The dynamic feature vector and the rule confidence vector are dimensionally aligned and concatenated to generate a fused input vector.

[0104] An adaptive weight is assigned to each vector using a gated neural network. The gated neural network consists of two fully connected layers: the first layer uses the ReLU activation function, and the second layer uses the Softmax function. The result is a dynamic feature vector weight (α) and a regular confidence vector weight (β), which satisfy α + β = 1.

[0105] Based on adaptive weights α and β, the dynamic feature vector and the rule confidence vector are weighted and fused to generate a fused feature vector.

[0106] The fused feature vector is input into the multi-task decision head for subsequent decision-making.

[0107] Alarm level generation: Outputs continuous alarm level values ​​through a fully connected layer.

[0108] Event type classification: The probability distribution of event types is output through a fully connected layer and the Softmax function.

[0109] Evidence summary generation: Based on the attention mechanism, key evidence is extracted from the fused feature vector to generate a structured text summary.

[0110] For evidence summary generation, the first step is to obtain the attention weights between the dynamic feature vector and the rule confidence vector. The top N features with the highest weights are selected as key evidence. Predefined templates, such as event type, confidence level, and evidence source, are then used to populate the key evidence, generating a structured text summary for further analysis and decision-making.

[0111] Based on the alarm level, event type probability distribution, and evidence summary output by the decision head, real-time security monitoring results are generated.

[0112] Real-time alert levels: Based on dynamically changing environmental data, alert levels are provided in real time, such as high risk or low risk.

[0113] Event type probability distribution: Determine the type of event and provide the corresponding probability distribution, such as fire, intrusion, etc.

[0114] Evidence summary outputs a structured summary of evidence, including information such as event type, confidence level, and source of evidence, to facilitate decision-making by security personnel.

[0115] The above embodiments provide a powerful intelligent security monitoring solution for smart parks. By combining multi-dimensional perception and efficient data processing technology with drones and ground sensors, it provides real-time decision support and early warning information for security management personnel.

[0116] Example 2: This example demonstrates the application of the present invention to the deployment of an intelligent three-dimensional security monitoring system in the XX Science and Technology Park. This science and technology park covers an area of ​​approximately 2 square kilometers and includes office areas, R&D buildings, production workshops, warehouses, an energy center, and public green spaces. The park experiences complex personnel flow, has high asset value, and imposes stringent security requirements.

[0117] First, a multi-dimensional, three-dimensional sensing network is constructed within the science and technology park:

[0118] The drone sensing network is deployed at three fixed take-off and landing points, employing a heterogeneous fleet of six drones of three different types. Two visible-light drones, equipped with 30x optical zoom cameras, are used for large-area patrols and detailed identification. Their cruising speed is 8 m / s, with a flight time of 30 minutes and a single-flight coverage area of ​​approximately 0.5 square kilometers. One infrared temperature measurement drone, equipped with an 80mm infrared lens and a temperature measurement accuracy of ±2℃, is used for nighttime patrols and equipment thermal anomaly detection. Its cruising speed is 6 m / s, with a flight time of 25 minutes. One laser scanning drone, equipped with a 250-line lidar and a point cloud density of 0.1 points / cm², is used for perimeter terrain modeling and illegal structure detection. Its flight speed is 5 m / s, with a flight time of 20 minutes.

[0119] The heterogeneous drone fleet collaborates with the ground control center in real time via 5G network, dynamically deploying resources based on preset patrol routes and real-time alarms. When an infrared drone detects an abnormal temperature in a transformer area, it immediately notifies a visible light drone to proceed to the area for image verification, while a laser drone simultaneously scans the surrounding environment.

[0120] A ground-based fixed sensor network is deployed along the park's perimeter, main roads, key entrances and exits, and building rooftops; a total of 50 multispectral video surveillance nodes are included, comprising 40 high-definition visible light cameras with a resolution of 4MP and 10 thermal imaging cameras with a temperature measurement range of -40℃ to 650℃. Some cameras possess starlight-level night vision capabilities. Coverage ensures no blind spots, with 200% redundancy in critical areas such as warehouses and data centers. Auxiliary sensors are also deployed, including vibration sensors along the perimeter wire fence and critical pipelines for detecting shearing and climbing, and water immersion sensors for detecting leaks.

[0121] For the mobile IoT sensing network, a security team of 25 people within the park was equipped with UWB / BLE dual-mode positioning tags. UWB, with an average error of 0.8 meters, provides sub-meter level high-precision positioning for accurate location of personnel during emergency calls; BLE is used for tracking personnel activity areas and issuing area alarms at medium range (10-30 meters). The tags report location data every 2 seconds.

[0122] All data generated by sensing devices, including video streams, infrared images, point cloud data, vibration signals, and personnel location coordinates, first undergoes preliminary processing at edge computing nodes: Video streams are used for moving target detection (e.g., people, vehicles), extracting metadata such as target trajectory and speed to reduce transmission bandwidth requirements. Infrared data is used for temperature anomaly region segmentation. Point cloud data undergoes point cloud dimensionality reduction and feature extraction. Sensor data is filtered and denoised. The resulting preprocessed multimodal heterogeneous data, including target trajectories, temperature anomaly coordinates, point cloud features, vibration intensity, and personnel locations, is packaged into a unified format data packet and sent to the central security monitoring platform.

[0123] One hundred key monitoring points within the park, such as camera locations, access control points, and sensor locations, are designated as nodes. Edge weights between nodes are based on historical event correlations, such as the frequency of simultaneous events between two points and their physical distance, and are then normalized. A dynamic adjacency matrix is ​​used to represent the node connections at the current moment.

[0124] In this embodiment, target trajectories, personnel flow speeds, and equipment vibration frequencies of each node over the past 5 minutes are selected as preprocessed dynamic time-series data to be input. Spatial features and event coordination patterns in adjacent areas are extracted using a spatiotemporal graph convolutional layer GCN variant. The time series is processed using a gated temporal convolutional layer LSTM variant to capture the temporal dependencies of event development and obtain a continuous action sequence of intrusion behavior. Then, a two-layer Transformer encoder is used to encode the fused spatiotemporal features, generating a 128-dimensional dynamic feature vector that represents the current overall dynamic situation of the park.

[0125] For Bayesian network construction, key security event nodes are first defined: "Intrusion," "Fire," "Equipment Failure," and "Personnel Gathering," along with related state nodes: "Access Control Abnormality," "Excessive Temperature," "Severe Vibration," and "Number of People in Area > Threshold." Based on expert knowledge, namely security regulations and equipment maintenance manuals, dependencies between nodes are defined. For example, "Access Control Abnormality" and "Number of People in Area > Threshold" both point to the "Intrusion" event, constructing a Bayesian network topology with 15 nodes. A hierarchical neural network structure is used to model the conditional probability table (CPT) for each non-root node. The first sub-network divides the parent node states into 2-3 groups based on logical correlation, such as temporal correlation and spatial proximity, encodes the parent node states in each group, and outputs an intermediate representation (32 dimensions). For the second fusion network, the intermediate representations of each group are concatenated, with a total dimension of 96, and input to a fully connected layer, outputting the state probability distribution of the node, such as a "normal" / "abnormal" binary distribution, or a multivariate distribution for specific fault types.

[0126] The preprocessed state event data—namely, the number of consecutive access control failures, the average infrared temperature of a certain area, the reading of a pipeline vibration sensor, and the UWB positioning display of the population density of a certain area—are mapped to the corresponding leaf nodes of a Bayesian network and encoded as embedding vectors with a dimension of 16. The leaf node identifiers and embedding vectors are used as evidence vectors. A differentiable variational inference algorithm is employed; in this embodiment, KL divergence minimization is used for approximate inference to calculate the posterior probability distribution of the target event nodes.

[0127] For binary state events, such as "intrusion," the probability value of the "abnormal" state in its posterior probability distribution is taken as the confidence level. For multivariate state events, such as the specific type of "equipment failure," the complete state probability vector is output as the confidence level vector. Finally, a rule confidence level vector of length 20 is generated, corresponding to 20 predefined security events and their confidence levels.

[0128] The 128-dimensional dynamic feature vector and the 20-dimensional rule confidence vector are aligned in dimension. The rule confidence vector is copied 6 times or increased to 128 dimensions through linear transformation. Then, they are concatenated into a 256-dimensional fusion input vector.

[0129] A gated neural network with fused input vectors. This network consists of two fully connected layers:

[0130] First layer: 256 inputs -> 64 outputs, activation function ReLU.

[0131] The second layer consists of 64 inputs to 2 outputs (α, β), with a softmax activation function to ensure α + β = 1. For example, if the current input results in a network output of α = 0.7 and β = 0.3, it indicates that the dynamic feature vector contributes more.

[0132] Weighted fusion: Calculate the fusion feature vector = 0.7 Dynamic feature vector +0.3 Rule confidence vector (aligned dimensions).

[0133] The fused 128-dimensional feature vector is then input into the multi-task decision head:

[0134] A fully connected layer (128->1) outputs a sequential alarm level value between 0 and 10. The alarm level generation branch outputs 7.2, indicating a high current security risk.

[0135] A fully connected layer (128->20) is connected to a Softmax function, outputting a 20-dimensional event type probability distribution vector. Event classification branches are generated, [0.01, 0.02, 0.85, 0.05, ..., 0.01], representing the "intrusion" event with the highest probability (85%).

[0136] The evidence summary generation branch requires obtaining attention weights and extracting key evidence. An attention mechanism is designed to take the fused feature vector as input and calculate its similarity to the original dynamic feature vector and the original rule confidence vector, resulting in their respective attention weight vectors (dimensions 128 and 20). Then, for the attention weights of the dynamic feature vector, the features corresponding to the top 5 dimensions with the highest weights are selected; for the attention weights of the rule confidence vector, the confidence values ​​corresponding to the top 3 events with the highest weights are selected. These features / events constitute the key evidence.

[0137] Text generated from a predefined template: Event type: Intrusion, Confidence level: 85%.

[0138] Key evidence:

[0139] Source of evidence: Visible light from a drone; Content of evidence: Suspicious individuals were observed climbing over the wall.

[0140] Source of evidence: UWB positioning; Evidence content: Abnormally high population density in the walled area.

[0141] Source of evidence: Access control system; content of evidence: Access control on one side failed 3 times in a row.

[0142] Source of evidence: Spatiotemporal pattern; Evidence content: High incidence of intrusions in this area recently.

[0143] Source of evidence: thermal imaging; content of evidence: a moving heat source was observed outside the wall.

[0144] The summary visually presents the reasons and basis for the alert.

[0145] System performance: After deployment in XX Science and Technology Park, the system successfully alerted 5 real intrusion events during a 3-month test, with the average response time reduced to 45 seconds; the false alarm rate was reduced by 60% compared to traditional systems, effectively reducing the unnecessary deployment of security personnel; the generated structured evidence summary helped security supervisors quickly understand the full picture of the event, improving decision-making efficiency by 40%.

[0146] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0147] The above embodiments are merely illustrative examples of the present invention and do not constitute a limitation on the scope of protection of the present invention. Any designs that are the same as or similar to the present invention are within the scope of protection of the present invention.

Claims

1. A method for intelligent three-dimensional security monitoring in smart parks, characterized in that, Includes the following steps: A multi-dimensional three-dimensional perception network is constructed by utilizing UAV perception networks, ground-based fixed sensor networks, and mobile IoT perception networks. Based on the three-dimensional perception network, multimodal heterogeneous data is acquired and preprocessed. A security monitoring model is constructed by parallelizing spatiotemporal graph neural networks and Bayesian networks and adaptively fusing them at the decision layer. The preprocessed multimodal heterogeneous data is input into the security monitoring model. Dynamic temporal data within the preprocessed multimodal heterogeneous data is processed using a spatiotemporal graph neural network to obtain dynamic feature vectors. State event data is processed using a differentiable conditional probability model. Rule reasoning is performed based on a conditional probability table generated from expert rules, and a rule confidence vector is output. The process of processing the dynamic temporal data within the preprocessed multimodal heterogeneous data using a spatiotemporal graph neural network to obtain dynamic feature vectors includes the following steps: using key points in the monitoring area as nodes, spatiotemporal relationships between nodes as edges, and constructing a spatiotemporal graph from the nodes, edges, and adjacency matrix obtained from the dynamic temporal data; the spatiotemporal graph convolutional layer in the spatiotemporal graph neural network extracts spatial features from the spatiotemporal graph, and a gated temporal convolutional layer obtains spatiotemporal features; the encoder of the spatiotemporal graph neural network encodes the spatiotemporal features to generate dynamic feature vectors. The process of processing state event data using a differentiable conditional probability model, performing rule inference based on a conditional probability table generated from expert base rules, and outputting a rule confidence vector includes the following steps: treating security-related events as nodes in a Bayesian network and using probability dependencies defined by expert base rules as edges to construct a Bayesian network topology; constructing a differentiable conditional probability model based on the network topology defined by expert base rules and an initial conditional probability table, wherein the conditional probability table for each non-root node is obtained through neural network parameterization; mapping event and state record data in the state event data to corresponding leaf nodes in the Bayesian network topology, and encoding the observed data of the events and states into embedding vectors for the corresponding leaf nodes; using the corresponding leaf nodes and their embedding vectors as evidence vectors; performing probabilistic inference using the evidence vectors and the differentiable conditional probability model to obtain the posterior probability distribution of the target event node; and extracting the confidence of the target event based on the posterior probability distribution to generate a rule confidence vector. An adaptive gating fusion mechanism is designed to adaptively weight and fuse the dynamic feature vector and the rule confidence vector to generate multi-level decision outputs, including real-time alarm levels, event type probability distributions, and evidence summaries, as security monitoring results. The gating fusion mechanism includes: dimensional alignment and concatenation of the dynamic feature vector and the rule confidence vector to generate a fusion input vector; generating adaptive weights for the dynamic feature vector and the rule confidence vector through a gating neural network; wherein the gating neural network structure includes two fully connected layers, the first fully connected layer using the ReLU activation function and the second fully connected layer using the Softmax function; the adaptive weights include a dynamic feature vector weight α and a rule confidence vector weight β, satisfying α + β = 1; based on the adaptive weights, the dynamic feature vector and the rule confidence vector are weighted and fused to generate a fused feature vector; a multi-task decision head is introduced, and the fused feature vector is input into the multi-task decision head to generate multi-level decision outputs.

2. The intelligent three-dimensional security monitoring method for smart parks as described in claim 1, characterized in that, The construction of the multi-dimensional stereo perception network includes the following steps: The drone sensing network adopts a heterogeneous cluster collaborative architecture, including visible light drones, infrared temperature measurement drones, and laser scanning drones; The ground-based fixed sensor network deploys multispectral video monitoring nodes; The mobile IoT sensing network tracks personnel using UWB and BLE dual-mode positioning tags.

3. The intelligent three-dimensional security monitoring method for smart parks as described in claim 1, characterized in that, For each non-root node, a conditional probability table is modeled using a hierarchical neural network structure, including the following steps: The first layer of the sub-network groups the parent node states according to logical correlation and outputs the intermediate representation of each group of parent nodes; The second-layer fusion network takes the intermediate representations as input and outputs the state probability distribution of the node. The probabilistic reasoning employs a differentiable approximate reasoning algorithm.

4. The intelligent three-dimensional security monitoring method for smart parks as described in claim 1, characterized in that, The method for generating the rule confidence vector is as follows: The target event node is a binary state, and its positive class state probability is taken as the confidence level. The target event node is a multi-state entity, and the confidence level of its complete state probability vector is taken.

5. The intelligent three-dimensional security monitoring method for smart parks as described in claim 1, characterized in that, The multi-task decision head includes: The continuously generated alarm level values ​​are output through a fully connected layer and used as the alarm level generation branch. The event type probability distribution is output through a fully connected layer and a Softmax function, serving as the event classification branch. Based on attention weights, key evidence is extracted from the dynamic feature vector and rule confidence vector to generate a structured text summary, which serves as the evidence summary generation branch. The evidence digest generation includes: Obtain the attention weights of the fused feature vector, the dynamic feature vector, and the rule confidence vector; The top N features with the highest attention weights are selected as key evidence. A structured text summary is generated by filling key evidence into a predefined template; wherein the predefined template includes an event type field, a confidence level field, and an evidence source field.

6. An intelligent three-dimensional security monitoring system for smart parks, characterized in that, include: The data feature acquisition module is used to construct a multi-dimensional three-dimensional perception network using UAV perception network, ground fixed sensor network and mobile IoT perception network, and to acquire multimodal heterogeneous data and preprocess it based on the three-dimensional perception network. The monitoring data analysis module is used to combine spatiotemporal graph neural networks and Bayesian networks in parallel and adaptively fuse them at the decision level to build a security monitoring model; The preprocessed multimodal heterogeneous data is input into the security monitoring model. Dynamic temporal data within the preprocessed multimodal heterogeneous data is processed using a spatiotemporal graph neural network to obtain dynamic feature vectors. State event data is processed using a differentiable conditional probability model. Rule reasoning is performed based on a conditional probability table generated from expert rules, outputting a rule confidence vector. The process of processing the dynamic temporal data within the preprocessed multimodal heterogeneous data using a spatiotemporal graph neural network to obtain dynamic feature vectors includes the following steps: using key points in the monitoring area as nodes, spatiotemporal relationships between nodes as edges, and constructing a spatiotemporal graph from the nodes, edges, and adjacency matrix obtained from the dynamic temporal data; the spatiotemporal graph convolutional layer in the spatiotemporal graph neural network extracts spatial features from the spatiotemporal graph, and a gated temporal convolutional layer obtains spatiotemporal features; the encoder of the spatiotemporal graph neural network encodes the spatiotemporal features to generate dynamic feature vectors; the process of processing the dynamic temporal data using a differentiable conditional probability model... The probability model processes state event data, performs rule inference based on a conditional probability table generated from expert base rules, and outputs a rule confidence vector. The steps include: constructing a Bayesian network topology by treating security-related events as nodes and using probabilistic dependencies defined by expert base rules as edges; constructing a differentiable conditional probability model based on the network topology and initial conditional probability table defined by expert base rules, where the conditional probability table for each non-root node is obtained through neural network parameterization; mapping event and state record data in the state event data to corresponding leaf nodes in the Bayesian network topology, and encoding the observed data of the events and states into embedding vectors for the corresponding leaf nodes; using the corresponding leaf nodes and their embedding vectors as evidence vectors; performing probabilistic inference using the evidence vectors and the differentiable conditional probability model to obtain the posterior probability distribution of the target event node; and extracting the confidence of the target event based on the posterior probability distribution to generate a rule confidence vector. The security monitoring result acquisition module is used to design an adaptive gating fusion mechanism for decision-making levels. It adaptively weights and fuses the dynamic feature vector and the rule confidence vector to generate multi-level decision outputs, including real-time alarm levels, event type probability distributions, and evidence summaries, as security monitoring results. The gating fusion mechanism includes: dimensional alignment and concatenation of the dynamic feature vector and the rule confidence vector to generate a fusion input vector; generating adaptive weights for the dynamic feature vector and the rule confidence vector through a gating neural network; wherein the structure of the gating neural network includes two fully connected layers, the first fully connected layer using the ReLU activation function and the second fully connected layer using the Softmax function; the adaptive weights include a dynamic feature vector weight α and a rule confidence vector weight β, satisfying α + β = 1; based on the adaptive weights, the dynamic feature vector and the rule confidence vector are weighted and fused to generate a fused feature vector; and a multi-task decision head is introduced, into which the fused feature vector is input to generate multi-level decision outputs.

7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 5.

8. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 5.

Images