Security monitoring alarm linkage processing method and system

Abstract

The invention relates to the technical field of security and protection monitoring, and discloses a security and protection monitoring alarm linkage processing method and system. The method comprises the following steps: constructing a three-dimensional scene graph of a monitoring area; obtaining visual position information and wireless positioning information of a target, and executing constrained Kalman filtering fusion to obtain a target position; mapping the target position to the three-dimensional scene graph for threat assessment to obtain an intelligent alarm signal; and performing linkage control in the three-dimensional scene graph based on the intelligent alarm signal to obtain an equipment activation instruction and a blocking strategy. According to the method, high-precision three-dimensional target positioning is realized in a complex environment, an alarm decision strategy can be dynamically optimized according to historical statistical data, the problem of high false alarm rate caused by a fixed threshold mode is effectively reduced, and the stability and the reliability in different environment conditions are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the security monitoring technical field, and particularly relates to a security monitoring alarm linkage processing method and system. BACKGROUND

[0002] The traditional security monitoring system relies on a single camera sensor to detect and locate targets, and lacks the ability to deeply understand the three-dimensional spatial structure of the monitoring area. The existing security monitoring system adopts a planar monitoring layout and a static area division method, which cannot effectively handle the cross-space monitoring requirements in multi-story and complex building structures, resulting in the long-term existence of monitoring blind spots and insufficient positioning accuracy. It cannot dynamically adjust the alarm strategy according to the safety risks of different time periods and different areas, and lacks the ability to deeply analyze the behavior patterns of targets, which seriously affects the practicality and reliability of the security system. SUMMARY

[0003] The main purpose of the present application is to provide a security monitoring alarm linkage processing method and system. The present application realizes high-precision three-dimensional target positioning in complex environments, can dynamically optimize alarm decision strategies based on historical statistical data, effectively reduces the high false alarm rate caused by fixed threshold methods, and improves the stability and reliability in different environmental conditions.

[0004] To achieve the above purpose, the present application provides a security monitoring alarm linkage processing method, comprising the following steps: Constructing a three-dimensional scene graph of the monitoring area; Obtaining visual position information and wireless positioning information of the target, and performing constraint Kalman filtering fusion to obtain the target position; Mapping the target position to the three-dimensional scene graph for threat assessment to obtain an intelligent alarm signal; Based on the intelligent alarm signal, performing linkage control in the three-dimensional scene graph to obtain a device activation instruction and a blocking strategy.

[0005] Optionally, in the first implementation manner of the first aspect of the present application, the constructing a three-dimensional scene graph of the monitoring area comprises: Spatially dividing the monitoring area to obtain a set of spatial nodes; Calculating the dynamic safety level value of each spatial node according to the distance of each spatial node in the set of spatial nodes to the key area, the time period danger coefficient and the historical risk score; Based on the spatial position relationship of the set of spatial nodes, performing connectivity analysis to obtain a node topology connection relationship; Hierarchical semantic organization is performed on the set of spatial nodes, the node topology connection relationship and the dynamic safety level value to obtain a three-dimensional scene graph.

[0006] Optionally, in a second implementation form of the first aspect of the present application, the acquiring the visual position information and the wireless positioning information of the target and performing constraint Kalman filtering fusion to obtain the target position comprises: inputting the camera captured image into the YOLO-SecNet neural network for multi-task learning detection to obtain camera target detection data; performing pixel coordinate to world coordinate conversion from the bounding box coordinates in the camera target detection data to obtain visual position information of the target in the monitoring scene; performing time difference arrival measurement on the target signal through the ultra-wideband base station to obtain ultra-wideband ranging data; performing non-line-of-sight error compensation and three-dimensional positioning solution on the ultra-wideband ranging data to obtain wireless positioning information; performing constraint Kalman filtering fusion on the visual position information and the wireless positioning information to obtain the target position.

[0007] Optionally, in a third implementation form of the first aspect of the present application, the performing constraint Kalman filtering fusion on the visual position information and the wireless positioning information to obtain the target position comprises: setting the visual position information and the wireless positioning information as camera observation and ultra-wideband observation respectively and constructing a state vector; performing a prediction step and an update step of Kalman filtering based on the state vector to obtain a filtering position result; applying a physical constraint condition of speed limit and height boundary to the filtering position result to obtain a feasible domain constraint range; performing sequence quadratic programming optimization solution on the filtering position result within the feasible domain constraint range to obtain the target position.

[0008] Optionally, in a fourth implementation form of the first aspect of the present application, the performing sequence quadratic programming optimization solution on the filtering position result within the feasible domain constraint range to obtain the target position comprises: constructing a constraint minimization problem according to the filtering position result and a precision weighted matrix; constructing an augmented optimization function by combining an objective function of the constraint minimization problem with the feasible domain constraint range; performing gradient and Hessian matrix derivation on the augmented optimization function to obtain a search direction vector; determining an optimal iteration step length through the Armijo line search criterion on the search direction vector to obtain the target position.

[0009] Optionally, in a fifth implementation form of the first aspect of the present application, the mapping the target position to the three-dimensional scene graph for threat assessment to obtain an intelligent alarm signal comprises: constructing a trajectory feature vector according to the target position and historical motion data; matching a corresponding space node in the three-dimensional scene graph based on the target position and extracting a dynamic security level value of a neighborhood node to obtain a scene context feature; inputting the trajectory feature vector and the scene context feature into a threat assessment neural network for spatio-temporal feature fusion to obtain a probability distribution result; performing numerical calculation on the probability distribution result according to a threat level weight to obtain a target threat level; performing adaptive threshold judgment according to the target threat level to obtain an intelligent alarm signal.

[0010] Optionally, in the sixth implementation manner of the first aspect of the present application, the adaptive threshold judgment according to the target threat level to obtain an intelligent alarm signal comprises: performing weighted fusion of the target threat level, a dynamic security level value of a corresponding space node and a trajectory behavior feature to obtain an alarm probability; performing likelihood probability calculation of a normal behavior mode on the target motion trajectory through a hidden Markov model to obtain an abnormality index; calculating a reward signal according to statistical data of false alarm events, missed alarm events and accurate alarm events; inputting the reward signal, the alarm probability and the abnormality index into a deep Q network for threshold decision learning to obtain an intelligent alarm signal.

[0011] Optionally, in the seventh implementation manner of the first aspect of the present application, the linkage control based on the intelligent alarm signal in the three-dimensional scene graph to obtain a device activation instruction and a blocking strategy comprises: performing PID tracking control based on pixel error on a PTZ camera in response to the intelligent alarm signal to obtain a first-level response instruction of target locking; running a Dijkstra algorithm in the three-dimensional scene graph to calculate a graph distance of each space node to a target path to obtain a device activation probability value; screening a monitoring device that needs to be activated and identifying a key passage node to perform a blocking operation according to the device activation probability value and a preset threshold condition to obtain a second-level response control sequence and a third-level response control sequence; performing priority sorting on the first-level response instruction, the second-level response control sequence and the third-level response control sequence to obtain a device activation instruction and a blocking strategy.

[0012] Optionally, in an eighth implementation form of the first aspect of the present application, the running Dijkstra algorithm in the three-dimensional scene graph to calculate the graph distance of each space node to the target path to obtain the device activation probability value comprises: performing trajectory prediction on the target current motion state to obtain a predicted motion trajectory; performing space node matching positioning of each position point in the predicted motion trajectory in the three-dimensional scene graph to obtain a key space node sequence; performing Dijkstra algorithm traversal calculation using the key space node sequence as a starting point set and using node topology connection relationship to obtain the graph distance of each monitoring device node in the scene to the predicted path; performing probability conversion based on the graph distance using an exponential decay function to obtain the device activation probability value.

[0013] The present application also provides a security monitoring alarm linkage processing system, comprising: a construction module configured to construct a three-dimensional scene graph of a monitoring area; a filter fusion module configured to obtain visual position information and wireless positioning information of a target and perform constrained Kalman filter fusion to obtain a target position; a threat assessment module configured to map the target position to the three-dimensional scene graph to perform threat assessment and obtain an intelligent alarm signal; a linkage control module configured to perform linkage control in the three-dimensional scene graph based on the intelligent alarm signal to obtain a device activation instruction and a block strategy.

[0014] In summary, the technical scheme provided by the present application has the understanding ability of the complete three-dimensional space structure of the monitoring area by constructing a three-dimensional scene graph comprising space nodes, topological connection relationships and dynamic security levels, can accurately identify the spatial correlation relationship across floors and regions, and effectively solves the spatial cognition limitation problem of the traditional plane monitoring system. The dual constraint mechanism of camera visual information and ultra-wideband wireless positioning information is adopted, the systematic errors of a single sensor are eliminated through constraint Kalman filtering fusion and physical constraint optimization, high-precision three-dimensional target positioning is realized in a complex environment, and the technical bottleneck of insufficient positioning accuracy of the traditional single sensor is overcome. The spatial semantic information of the three-dimensional scene graph and the target motion trajectory characteristics are combined, the spatio-temporal feature fusion analysis is performed through the threat evaluation neural network, not only the target behavior mode is considered, but also the spatial context information is fully utilized, and the accuracy and explainability of the abnormal behavior recognition are improved. The normal behavior baseline is established through the hidden Markov model, and the adaptive threshold adjustment mechanism of deep reinforcement learning is adopted, the alarm decision strategy can be dynamically optimized according to historical statistical data, the high false alarm rate problem caused by the fixed threshold mode is effectively reduced, and the high detection rate of real threats is maintained. The layered linkage control mechanism based on the three-dimensional scene graph is adopted, the path planning and device activation probability calculation are performed through the Dijkstra algorithm, accurate monitoring device selection and time-sequenced response control are realized, the resource waste of the traditional blind activation mode is avoided, and the pertinence and timeliness of the linkage response are greatly improved. Through the physical constraint optimization of sequential quadratic programming and the non-line-of-sight error compensation mechanism, the multipath propagation and shielding interference of the complex indoor environment can be adapted, and the stability and reliability in different environmental conditions are improved. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a security monitoring alarm linkage processing method step schematic diagram in an embodiment of the present application; Figure 2 is a security monitoring alarm linkage processing system structure block diagram in an embodiment of the present application.

[0016] The implementation, functional characteristics and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0018] With reference to Figure 1 , the present embodiment provides a security monitoring alarm linkage processing method, comprising the following steps: S1, construct a three-dimensional scene graph of the monitoring area; Specifically, the entire monitoring area is spatially divided, and by combining building structure diagrams, floor design diagrams and functional partition planning data, the monitoring space is divided into a plurality of space nodes with unique identifiers according to the floor structure and the functional area boundary. Each space node has a clear three-dimensional coordinate attribute, floor number, area type code and other basic elements, forming a space node set. The dynamic security level value of each node in the space node set is calculated respectively. The dynamic security level value is based on the joint modeling of three input factors: the spatial Euclidean distance from the node to the key area (such as the entrance, intersection, key area, etc.); the risk coefficient in the corresponding period, which is divided into time slices according to the day and night time sequence, personnel density distribution model or security level rules to obtain the value; the node-level risk score established based on historical event records, reflecting the frequency and severity of abnormal or illegal events in the past data. The three input factors are combined by an exponential decay function and a weighted linear combination to form a dynamic security level index, which real-time expresses the safety sensitivity of each node in the current scene. According to the geometric position relationship of all space nodes in the three-dimensional coordinate system, the horizontal adjacency, vertical accessibility and cross-regional structure connection relationship are analyzed, the topological connection matrix between space nodes is established, and the topological connection matrix reflects the physical path accessibility and constitutes the edge set in the graph structure. The space node set, the topological connection relationship between nodes and the corresponding dynamic security level value are fused, and based on the tree structure model of floor-function block, hierarchical semantic organization is carried out, and in the graph structure, semantic fragments are formed in layers and extended attributes such as area function label, passing permission level, number of people that can be accommodated are injected, to generate a three-dimensional scene graph with geometric, topological and semantic triple expression ability.

[0019] S2, obtain the visual position information and the wireless positioning information of the target, and perform constraint Kalman filtering fusion to obtain the position of the target; Specifically, the continuous image frames collected by the camera are input into a YOLO-SecNet neural network model built on the YOLOv5 framework. The network adopts a multi-task learning architecture, adds an abnormal behavior recognition branch on the basis of performing regular target bounding box and class detection, so that the output result contains bounding box coordinates, class confidence and behavior feature vector and other information. The center pixel coordinates of the bounding box are extracted, and the preset camera internal and external parameter matrix is combined to use the perspective projection inverse transformation and geometric correction method to complete the mapping of the pixel coordinates to the world coordinate system, and obtain the three-dimensional visual position information of the target in the real monitoring scene. At the same time, multiple groups of ultra-wideband base stations deployed in the monitoring area listen to the wireless pulse signals emitted by the target wearing label, use the time difference arrival algorithm to measure the time difference of the signals received by different base stations, and establish an over-determined equation set based on the known spatial coordinates of the base stations to solve the three-dimensional ranging. In order to correct the non-line-of-sight error caused by obstacles, walls and other factors, an error compensation mechanism based on a signal-to-noise ratio attenuation model is introduced to weight and correct the original ranging data, and the wireless positioning accuracy is improved. The visual position information and the wireless positioning information optimized after compensation are input into an extended Kalman filter framework, a high-dimensional state vector containing the target position, velocity and behavior state is defined, and the camera observation and UWB observation are fused in the prediction-update cycle respectively, the weight is adjusted dynamically combined with the observation error covariance of each observation, and the feasible solution space is constructed by applying physical constraint conditions such as velocity upper limit, height boundary and behavior rationality. Through the constraint projection algorithm, the intermediate estimated state is projected and corrected in each update step to obtain the target position.

[0020] S3, mapping the target position to the three-dimensional scene graph for threat assessment to obtain an intelligent alarm signal; Specifically, based on the target position and the historical position sequence, a trajectory feature vector containing a time-space state is constructed, which includes the three-dimensional position coordinate change of the target, the velocity vector, the behavior feature code, and the trajectory continuity at different time points, forming a space-time representation reflecting the motion trend and behavior pattern. In the three-dimensional space graph, the space node matching the current target position is retrieved, and based on the topological connection relationship of the space node, the neighborhood set composed of its adjacent nodes is extracted, and the scene context features are formed by reading the dynamic security level values corresponding to the neighborhood nodes, representing the risk environment state of the target area and its surrounding area. At the same time, the risk coefficient of the current time period and the historical risk weight distribution are considered to enhance the semantic dimension. The trajectory feature vector and the scene context feature are input into the threat assessment neural network model, which internally uses embedding layer, position encoding module and multi-layer time sequence Transformer structure to jointly model the two types of features. Through the self-attention mechanism, the correlation between the trajectory state and the space semantics is extracted, and a multi-class distribution vector containing the probability of different threat levels is output. The probability distribution result is weighted with the preset threat level weight coefficient to convert it into a quantifiable target threat level value, expressing the comprehensive risk level of the current target in the environment. According to the characteristics of the current period, the historical false alarm rate statistical results and the real-time threat level trend, the threshold reference curve is adaptively set, and the target threat level value is dynamically determined. When the target threat level value exceeds the set threshold, it is considered that the abnormal behavior trigger condition is established, and an intelligent alarm signal is output.

[0021] S4, based on the intelligent alarm signal, perform linkage control in the three-dimensional scene graph to obtain device activation instructions and blocking strategies.

[0022] Specifically, in response to the intelligent alarm signal trigger, a first-level response task is performed, that is, a PTZ camera with a pan-tilt rotation capability is called to perform pixel-level locking control on the current location of the target. The pixel deviation is input into a PID controller for proportional, integral, and differential calculation to obtain a real-time pan-tilt rotation control amount, which is used to adjust the direction of the camera so that the target is continuously maintained in the center area of the screen. A first-level response instruction is generated to drive the target tracking module. In the three-dimensional scene graph that has been constructed, the Dijkstra shortest path algorithm is used to calculate the graph distance of all other space nodes in the scene graph relative to the target node, which reflects the physical accessibility between nodes. The node connectivity and the passing weight are considered, and the graph distance is mapped to a device activation probability value through an exponential decay function. The greater the value, the higher the relevance of the device to the target. On this basis, the device activation probability is screened according to the preset activation threshold to select the monitoring devices that need to be activated to form a second-level response control sequence. At the same time, the channel connectivity of all path nodes in the graph is analyzed to identify the channel nodes that can pass to high-risk areas or key target areas, and the channel nodes are executed to build a defense boundary to generate a third-level response control sequence. The target locking instruction of the first-level response, the device activation sequence of the second-level response, and the area locking control instruction of the third-level response are collected and arranged, and the comprehensive priority value is calculated according to the response urgency, target dynamic change rate, and device control delay. All instructions are prioritized to form an ordered device activation instruction list and area locking strategy plan.

[0023] In one example, a three-dimensional scene graph of a monitoring area is constructed, including: spatially dividing the monitoring area to obtain a set of space nodes; calculating a dynamic security level value of each space node in the set of space nodes according to the distance of each space node to a key area, a time period danger coefficient, and a historical risk score; performing connectivity analysis based on the spatial position relationship of the set of space nodes to obtain a node topology connection relationship; performing hierarchical semantic organization on the set of space nodes, the node topology connection relationship, and the dynamic security level value to obtain a three-dimensional scene graph.

[0024] In this example, based on the physical structure, building plan and function distribution data of the monitoring area, the overall space is structurally parsed, the continuous space area is divided into discrete sub-areas with boundary definition by adopting vertical and horizontal combined division method, and a unique identifier is set in each sub-area to generate a set of space nodes, wherein each space node records three-dimensional coordinate information, floor number, function area type code and auxiliary positioning parameter, constitutes a node database with space attributes and semantic labels, and the indoor building unit is accurately modeled by space division, and the area purpose (such as entrance, corridor, stairwell, guard area, key room, etc.) is semantically classified. In order to give the space node a dynamic perceptible safety attribute, a dynamic evaluation mechanism of node-level safety level value is constructed, and three types of key factors are considered: the spatial distance of each space node to one or more key areas, measured by Euclidean distance, and the key area weight (such as the main entrance weight is higher than that of ordinary passage); the danger coefficient corresponding to the current time period, based on the personnel density statistical law within 24 hours, the event high incidence time period or the alarm level standard, different time intervals are divided and valued respectively; the historical risk score of the node, by analyzing the historical alarm data, behavior anomaly record and event frequency to extract statistical characteristics, forming a static score reflecting the safety risk potential of the area itself. After the above three factors are combined by exponential decay function and linear weighting, the dynamic safety level value of each space node is obtained, reflecting the risk state of the space node under different context conditions. Based on the geometric position relationship, physical accessibility and functional connectivity between nodes, the topological structure is modeled, the connection relationship between the passable paths (such as doors, stairs, passages) is established, the topological edge set between the space nodes is constructed, and the direct connection and indirect path between the nodes are expressed in the form of graph structure, for example, when there is a stair passage or adjacent door connection between two space nodes, an edge is added in the graph structure and its physical length, passing direction or impedance weight is marked, to support path planning, blocking analysis and activated area propagation model. On the basis of node attributes, connectivity structure and safety level, the space nodes are vertically layered according to the floor structure and horizontally semantically organized according to the area function and structure layout in each layer, and the hierarchical logic and spatial context are described by tree or graph structure. Through the hierarchical semantic organization process, a three-dimensional scene graph model is formed, which has geometric space structure, functional semantic label, topological accessibility and spatio-temporal safety state.

[0025] In one example, the visual position information and the wireless positioning information of the target are acquired, and a constrained Kalman filter fusion is performed to obtain the position of the target, including: The camera-acquired image is input into the YOLO-SecNet neural network for multi-task learning detection to obtain camera target detection data; Converting the boundary box coordinates in the camera target detection data from pixel coordinates to world coordinates to obtain visual position information of the target in the monitoring scene; Measuring the time difference of arrival of the target signal by the ultra-wideband base station to obtain ultra-wideband ranging data; Compensating the non-line-of-sight error of the ultra-wideband ranging data and solving the three-dimensional positioning to obtain wireless positioning information; Constraining Kalman filtering fusion of the visual position information and the wireless positioning information to obtain the target position.

[0026] In this example, the images collected in real time by the cameras deployed in the monitoring area are input into the YOLO-SecNet neural network model, which is improved on the basis of the YOLOv5 infrastructure and fuses the dual-task structure of target detection and behavior recognition. Its multi-task learning framework can simultaneously output multi-dimensional detection results containing bounding box coordinates, class labels, confidence scores, and abnormal behavior feature vectors. Through the sharing of feature extraction networks and the decoupling structure of branch tasks, the detection and recognition accuracy and timeliness are improved. The center point pixel coordinates of the bounding box are extracted from the camera target detection data, and the pixel coordinates are converted into world coordinates through the space conversion operation of the camera's intrinsic matrix (containing focal length, principal point coordinates, etc.) and extrinsic matrix (representing the position and orientation of the camera in the scene coordinate system). The method of inverse projection transformation and geometric constraint solution is used to construct the space ray in the camera projection model and the intersection of the assumed ground plane or structure point to realize depth recovery, and output the three-dimensional visual position information of the target in the world coordinate system of the monitoring scene. At the same time, the ultra-wideband (UWB) positioning channel is started, and the UWB tag carried on the target emits high-frequency pulse signals which are listened to by multiple UWB base stations deployed at known positions. The time difference between the base stations is used for time difference arrival measurement, and a multi-lateral ranging equation set is constructed between the time difference information and the spatial coordinates of the target. The original ultra-wideband ranging data is obtained by solving the multi-lateral ranging equation set. In order to improve the reliability of positioning in complex environments, especially in the case of non-line-of-sight propagation error interference in indoor environments, an error compensation method based on a signal-to-noise ratio degradation model is introduced. An error correction coefficient related to the environment is introduced for each base station ranging value. The compensation model performs exponential function normalization operation on the actual measured signal-to-noise ratio value and the preset reference threshold, effectively eliminating systematic errors caused by physical mechanisms such as shielding and reflection. The compensated multi-base station ranging values are input into the three-dimensional positioning calculation module again, and the spatial positioning inverse calculation is completed under the premise of combining the base station geometric layout conditions to obtain the target wireless positioning information. The visual positioning result and the ultra-wideband positioning result are input into the extended constraint Kalman filter structure as observation data, the state vector containing the target three-dimensional position, velocity, acceleration and behavior variables is defined, and the state transition model and observation model are set. Through the filtering prediction and correction iteration process, the state is dynamically estimated. In the state update step, the feasible solution domain composed of physical and environmental constraint mechanisms is introduced, including position range boundary, maximum velocity limit, floor height level constraint and behavior value rationality limit. The predicted state is projected into the optimal approximate solution that satisfies the feasible domain, and the quadratic programming optimization method is used to realize high-precision correction. The target global position estimation result that meets the physical conditions, time sequence consistency and multi-source fusion accuracy optimization requirements is output.

[0027] The camera-acquired image is input into a YOLO-SecNet neural network for multi-task learning detection to obtain camera target detection data, including: performing multi-scale pyramid decomposition processing on the camera-acquired image to obtain a feature image sequence containing different resolution levels; performing adaptive weight allocation processing on each scale feature in the feature image sequence based on an attention mechanism to obtain enhanced feature representation highlighting a target region; performing hierarchical information integration processing on the enhanced feature representation through a cross-scale feature fusion network to obtain a fusion feature map that takes into account both detailed and global information; dynamically adjusting a region of interest extraction strategy of the detection network according to light changes and target scale distribution of a monitoring scene to obtain adaptive target candidate regions; inputting the fusion feature map and the target candidate regions into an improved detection head for multi-task joint prediction processing to obtain camera target detection data containing accurate bounding boxes, high confidence, and rich behavior features.

[0028] In one example, the visual position information and the wireless positioning information are fused by constrained Kalman filtering to obtain a target position, including: The visual position information and the wireless positioning information are set as camera observations and ultra-wideband observations respectively, and a state vector is constructed; A prediction step and an update step of Kalman filtering are performed based on the state vector to obtain a filtered position result; Physical constraint conditions of speed limitation and height boundary are applied to the filtered position result to obtain a feasible domain constraint range; The filtered position result is sequentially solved by quadratic programming optimization within the feasible domain constraint range to obtain the target position.

[0029] In this example, the visual position information is defined as camera observations, the three-dimensional wireless positioning information obtained by multi-base station ranging back calculation and error compensation is defined as ultra-wideband observations, and a unified state vector structure for fusion state estimation is constructed based thereon. The state vector includes position components (x, y, z) of the target in three-dimensional space, velocity components (vx, vy, vz), and additional behavior state variables for dynamic modeling, forming a high-dimensional, continuously differentiable state expression form. Then, the state transition model and the observation model of the system are constructed. In the state transition model, a uniform motion or acceleration driven model is used to describe the state evolution process of the target between consecutive frames, and Gaussian distributed process noise is introduced to reflect system disturbances and modeling errors. In the observation model, the observation matrix is set to project the state vector into the camera observation space and the UWB ranging space, and the observation noise covariance matrix is introduced to describe the error statistical characteristics of the two sensors, so as to realize system modeling under the premise of observation fusion. Based on this model, the Kalman filter performs the standard recursive process, that is, in the prediction step, the state prior estimate of the current frame is calculated using the previous frame state vector and the state transition matrix, and the covariance prior is also calculated to represent the estimation uncertainty; in the update step, the observations of the camera and the UWB are introduced respectively, and the state vector is updated by calculating the observation residual and the Kalman gain, and the covariance is updated, and the filtered intermediate state estimation result is output. The filtered position result combines the information of the two sensors, and the weight is adaptively allocated using the system dynamics and statistical characteristics. Considering the restriction conditions in the actual physical environment, the filtered position result is physically constrained, a feasible region defined by a multi-dimensional constraint boundary is constructed by setting the upper limit of the speed (for example, the horizontal direction is not more than 6 m / s, and the vertical direction is not more than 2 m / s) and the boundary conditions of the spatial position (for example, the height of the z-axis cannot exceed the boundary of the floor structure), and the feasible region describes the state range allowed by the target motion state under the real physical conditions. In order to map the state estimation value output by the Kalman filter into the feasible region and ensure the continuity and optimality of the solution, a sequential quadratic programming algorithm based on the minimization of the objective function is used for constraint correction. The algorithm takes the filtered output state as the original input, takes the feasible region as the constraint set, and takes the weighted norm of the state deviation as the objective function for optimization solution. The KKT condition system of the optimization problem is established by constructing the Lagrange function and introducing the Lagrange multiplier term, and the Newton iteration method is used to linearly approximate the second derivative structure of the objective function near the current search point, so as to realize fast convergence, and the optimal feasible state closest to the original filtered solution that satisfies all physical constraint conditions is output as the final target position output.

[0030] The prediction step and the update step of the Kalman filter based on the state vector are performed to obtain a filtering position result, including: establishing an adaptive filter set based on noise characteristics of camera observation and ultra-wideband observation to perform signal quality evaluation processing and obtain real-time reliability indicators of each sensor; dynamically adjusting an observation noise covariance matrix according to the reliability indicators to obtain an adaptive noise model reflecting a current performance state of the sensor; detecting and suppressing abnormal observation values through a robust Kalman filter algorithm to obtain a reliable observation data set eliminating wild value interference; performing variational Bayesian filter update processing based on the adaptive noise model and the reliable observation data set to obtain a state estimation result considering uncertainty propagation; and performing multi-step prediction consistency verification and error covariance adaptive adjustment processing on the state estimation result to obtain a filtering position result with robustness guarantee.

[0031] In one example, the filtering position result is subjected to sequence quadratic programming optimization solution within a feasible region constraint range to obtain a target position, including: A constrained minimization problem is constructed according to the filtering position result and a precision weighted matrix; An augmented optimization function is constructed by combining a target function of the constrained minimization problem with the feasible region constraint range; Gradient and Hessian matrix derivation are performed on the augmented optimization function to obtain a search direction vector; An optimal iteration step is determined for the search direction vector through Armijo line search criteria to obtain the target position.

[0032] In this example, the intermediate state estimation results output by the Kalman filtering process are taken as the initial reference position, which is modified under the constraint condition. A minimization problem is constructed to minimize the weighted distance between the current estimated position and the optimal feasible position, where the precision weighting matrix is used to reflect the uncertainty of each dimension of the state estimation, and the error penalty weight of the direction with high confidence is higher to reflect the contribution of different dimensions to the overall estimation reliability. The minimization problem is combined with the target physical constraint condition to construct an augmented optimization function containing physical boundary restrictions. The augmented function contains both the target deviation term and a set of inequality or equality constraint terms, which define the conditions that the target position must satisfy in space, such as speed upper limit, acceleration limit, floor height boundary, and behavior variable value range. To obtain an updated path that can make the augmented optimization function converge to the optimal solution, the gradient vector and Hessian matrix of the augmented function are derived to extract the function change direction and local curvature information at the current estimated point, and based on this, the optimal search direction vector from the current position to the next step is calculated. The optimal search direction can ensure that the optimal point is approached along the most effective path within the constraint range. Since directly moving along the search direction will cause the target function to oscillate or diverge, a step control mechanism is introduced to fine-tune the movement amplitude, and the Armijo line search criterion is used for adaptive step selection. The Armijo line search criterion evaluates whether the change of the target function under different step sizes meets the sufficient descent condition, gradually reduces the step size until the minimum descent amplitude requirement is met, and determines the optimal step length under the current search direction. By updating the current position along the search direction vector with the optimal step size, the process is iteratively executed until the search direction change rate tends to be stable or the change of the target function is less than the preset convergence threshold, and the current estimated position is taken as the final output target position.

[0033] In one example, the target position is mapped to a three-dimensional scene graph for threat assessment to obtain an intelligent alarm signal, including: Constructing a trajectory feature vector according to the target position and historical motion data; Matching the corresponding spatial node of the target position in the three-dimensional scene graph and extracting the dynamic security level value of the neighborhood node to obtain the scene context feature; Inputting the trajectory feature vector and the scene context feature into a threat assessment neural network for spatio-temporal feature fusion to obtain a probability distribution result; Numerical calculation of the probability distribution result according to the threat level weight to obtain the target threat level; Adaptive threshold judgment according to the target threat level to obtain an intelligent alarm signal.

[0034] In this example, the goal is to establish a representation of the spatial behavior of the target at the continuous moment in the monitoring area. By collecting historical motion data such as the target's position information, speed change, acceleration trend, and dwell time over a period of time, and combining it with the current spatial position to form a trajectory feature vector, the target's motion trend and state evolution are reflected, and potential signs of abnormal behavior in the trajectory are provided, such as frequent turns, long stays, or sudden speed changes. The current target position is mapped to the spatial nodes in the three-dimensional scene graph. The mapping process matches the three-dimensional coordinates of the target position with the spatial range boundaries of each node in the scene graph, identifies the main node where the target is located, and extracts the neighborhood node set in the graph structure according to the topological connection relationship of the scene graph, and reads the dynamic security level value of the neighborhood node at the current period. The dynamic security level value reflects the potential threat level of the target's surrounding space environment, and the resulting scene context features are supplemented as spatial semantic background information into the evaluation model. The trajectory feature vector and the scene context feature are input into the threat assessment neural network for joint processing. The threat assessment neural network uses a spatio-temporal fusion structure design, including a trajectory encoding layer, a position embedding layer, a self-attention mechanism module, and a semantic fusion layer. The trajectory feature vector embeds the trajectory state into a high-dimensional representation space to extract time evolution features. The scene context feature structures the spatial semantic structure and dynamic security level in the scene context through graph structure processing or graph neural network mechanism, and interacts the two types of features through the fusion layer, so that the model can simultaneously focus on the target behavior evolution path and the changing trend of its surrounding space environment. A multi-class probability distribution result is output, representing the probability value of the target belonging to each threat level in the current state. According to the threat level weight parameters set, the probability distribution components are weighted and added according to the weight, and the multi-class output is mapped to a continuous risk score value, which is the threat level of the current target under the comprehensive judgment of the spatio-temporal context, to measure the degree of discordance and potential security risk between the target's current behavior and the scene. An adaptive threshold determination mechanism is used to dynamically adjust the threshold according to the current environmental dynamic factors (such as time period, area type, personnel density), historical false alarm records, and typical behavior distribution rules, and to make real-time judgments combined with the threat level score calculated at the moment. When the score exceeds the current threshold, it is considered that the target constitutes a threat, and an intelligent alarm signal is generated. Otherwise, the monitoring state is maintained, and the behavior data is accumulated for the next evaluation.

[0035] The trajectory feature vector and the scene context feature are input into a threat assessment neural network for spatio-temporal feature fusion to obtain a probability distribution result, including: converting the trajectory feature vector into a time series graph structure and extracting the time dimension dependency to obtain a time series graph representation containing motion pattern evolution; constructing a spatial correlation graph neural network based on the spatial topology of a three-dimensional scene graph for information propagation processing between nodes to obtain enhanced node features that fuse the influence of spatial neighborhoods; performing joint modeling and processing of the time series graph representation and the enhanced node features through a spatio-temporal graph convolution network to obtain comprehensive feature representation that captures spatio-temporal coupling relationships; designing an adaptive feature fusion module based on a multi-head attention mechanism to perform weight allocation processing on multi-modal features to obtain a fusion feature vector that highlights key threat patterns; inputting the fusion feature vector into an incremental learning threat classifier for online model updating and prediction processing to obtain a threat level probability distribution result that adapts to environmental changes.

[0036] In one example, an adaptive threshold value is determined according to the target threat level to obtain an intelligent alarm signal, including: The target threat level is weighted and fused with the dynamic security level value of the corresponding spatial node and the trajectory behavior feature to obtain an alarm probability; The likelihood probability of the target motion trajectory is calculated through a hidden Markov model for a normal behavior pattern to obtain an abnormality index; A reward signal is calculated according to statistical data of false alarm events, missed alarm events, and accurate alarm events; The reward signal, the alarm probability, and the abnormality index are input into a deep Q network for threshold decision learning to obtain an intelligent alarm signal.

[0037] In this example, the target threat level, space environment state and behavior pattern three types of information are uniformly expressed and fused, the threat level value output by the threat assessment neural network is taken as the primary decision factor, and the dynamic security level of the target current space node in the three-dimensional scene graph is extracted synchronously. The dynamic security level value combines the distance from the node to the key area, the risk coefficient of the current period and the risk score calculated by the historical event frequency to reflect the spatial risk background of the environment where the target is located. The key behavior characteristics in the target trajectory such as speed change trend, residence time, direction offset and other parameters are taken as the behavior characteristic vector input. Through setting the fusion weight coefficient, the three are linearly or nonlinearly combined to obtain the alarm probability value of the target in the current scene state. At the same time, in order to improve the recognition ability of the system to the abnormality of the behavior sequence, a trajectory behavior evaluation method based on hidden Markov model is introduced, a state transition model and an observation output model of normal behavior are constructed, and the model parameters are trained based on historical normal trajectory samples, so that they can efficiently represent the trajectory evolution characteristics of the target in the abnormal state. On this basis, the motion trajectory of the current target is evaluated in terms of likelihood. The generation probability value of the current trajectory under the normal behavior model is calculated by the forward probability algorithm. The lower the likelihood probability value, the more the trajectory deviates from the normal mode. The abnormality index obtained by converting the likelihood probability value is taken as the second behavior risk evaluation data. Statistical analysis is performed on the historical alarm execution results. The number of false alarm events, missed alarm events and accurate alarm events triggered by the system in the past period of time is counted respectively. According to the resource waste caused by false alarm, the security risk that may be caused by missed alarm and the positive contribution of accurate alarm to the security system, a reward function is constructed by giving different weight factors to each type of event. The value of the reward function is used as the decision feedback signal in the reinforcement learning process to evaluate the overall performance of the alarm system under the current threshold strategy. The alarm probability, abnormality index and reward signal obtained by the above construction are jointly input into the Q network based on the deep reinforcement learning structure. The deep Q network maps the environment state to the alarm threshold adjustment action space, and estimates the long-term return of each possible action in the current state. Through the full connection layer, multi-batch experience replay and target network parameter update mechanism in the network structure, it is constantly trained and optimized, so that it can dynamically adjust the alarm threshold level under different scene states and trajectory behavior patterns to realize the intelligent alarm decision strategy. When the decision result output by the Q network shows that the current risk level has exceeded the optimal threshold point learned dynamically, the intelligent alarm signal is generated and pushed to the response linkage module.

[0038] The adaptive threshold determination according to the target threat level obtains an intelligent alarm signal, including: data integrity verification processing of the target threat level based on sensor data interference attack detection in the multi-region security monitoring network, to obtain adaptive adjustment variables containing threat level deviation and regional security control error; alarm trigger condition evaluation processing is performed according to the dynamic event trigger mechanism of the security performance standard dependence established by the adaptive adjustment variables, to obtain trigger threshold parameters that reduce unnecessary data transmission while ensuring the frequency stability and regional connection capability of the monitoring network; the constraint boundary calculation processing is performed by designing the security constraint set and the single-step response set respectively through the offline optimization problem, to obtain the terminal constraint boundary and the approximate single-step response range of the threat evaluation; the dynamic control strategy under the threat condition outside the constraint set is obtained by performing real-time decision processing based on the control law of the online optimization problem designed based on the boundary range of the security constraint set; the dynamic event trigger mechanism, the terminal constraint boundary and the dynamic control strategy are fused by the multi-step model predictive control algorithm, to obtain a multi-step prediction intelligent alarm signal based on the security performance standard dependence.

[0039] In one example, based on the intelligent alarm signal, a linkage control is performed in a three-dimensional scene graph to obtain a device activation instruction and a blocking strategy, including: Performing a PID tracking control based on pixel error on the PTZ camera in response to the intelligent alarm signal trigger to obtain a first-level response instruction of target locking; Running a Dijkstra algorithm in the three-dimensional scene graph to calculate the graph distance of each space node to the target path to obtain a device activation probability value; According to the device activation probability value and the preset threshold condition, screening the monitoring devices that need to be activated and identifying the key channel nodes to perform blocking operations to obtain a second-level response control sequence and a third-level response control sequence; Prioritizing the first-level response instruction, the second-level response control sequence and the third-level response control sequence to obtain the device activation instruction and the blocking strategy.

[0040] In this example, after detecting the intelligent alarm signal trigger, the target tracking module is started, and the first-level operation task is preferentially responded, that is, a tracking control instruction is issued to the PTZ camera installed at the junction of the current area or target path, the position pixel coordinates of the target in the camera picture are obtained in real time during this process, and the deviation between the position pixel position and the preset picture center reference point is taken as the control input. A PID controller is constructed by combining three feedback items of proportion, integral and differential, which are respectively used to process the current error, accumulated error and error change rate, to generate control quantities for controlling the horizontal rotation and vertical pitching of the camera holder, so that the camera angle is dynamically adjusted to keep the target always centered in consecutive frames, forming an image acquisition link. The adjustment action constitutes a first-level response instruction, realizing target locking and image following in the initial stage of alarm. At the same time, in the scene graph structure, the current spatial node of the target is taken as the source point, and the shortest graph path distance from the current spatial node of the target to all other nodes in the graph is calculated by Dijkstra algorithm. In the calculation process, the connectivity between nodes, edge weight, path passability and floor structure information are considered to ensure that the path reasoning conforms to the structure topology and takes into account the physical constraints, and the graph distance values of all spatial nodes relative to the current target position are obtained. Based on the graph distance data, an exponential decay strategy is used to map the graph distance to the device activation probability. The closer the distance, the higher the activation probability, and vice versa. The response probability field of the space correlation is formed. According to the device activation probability and the preset response threshold, the cameras, sensors or sound warning devices that meet the activation conditions are screened and classified, and a monitoring device set covering the potential motion path of the target is constructed. The second-level response control sequence is generated in order of activation probability, and the accessibility analysis and key identification of the channel nodes involved in the path are performed. The high-risk channel nodes leading to dangerous areas, entrances or densely populated areas are preferentially identified, and the blocking action plan is generated combined with the preset containment strategy, such as closing automatic doors, starting warning lights, adjusting the broadcast system, etc. The third-level response control sequence is constructed. All response actions are sorted according to the priority parameters. The first-level response instruction has the highest priority because it involves real-time tracking, and the second-level response control sequence has intermediate priority because it can expand the coverage range. The third-level response control sequence is conditionally triggered or low priority because it contains blocking actions and may interfere with normal passage. According to the above priority mechanism, the device activation instruction set and the blocking strategy plan are output and sent to each control execution unit.

[0041] The device activation probability value and the preset threshold condition are used to screen the monitoring device to be activated and identify the key channel node to perform the blocking operation, to obtain the secondary response control sequence and the tertiary response control sequence, including: performing boundary adaptive coverage analysis processing based on the three-dimensional scene graph monitoring device coverage grid integrity, to obtain a device deployment adaptive adjustment scheme containing boundary monitoring point insertion and coverage path optimization; inputting the current threat level information and the dynamic security level value into an environment information driven dynamic neural network to perform high risk area identification processing, to obtain the distribution of key threat areas that need to be blocked; performing monitoring coverage online re-planning processing according to the distribution of key threat areas and the existing device activation state, to obtain a dynamic device scheduling strategy for threat diffusion paths; performing a dead zone escape optimization algorithm processing on the monitoring blind area identified in the dynamic device scheduling strategy, to obtain a monitoring dead angle compensation scheme to eliminate blocking loopholes; performing double-layer coverage path planning integration processing on the device deployment adaptive adjustment scheme and the monitoring dead angle compensation scheme, to obtain the secondary response control sequence and the tertiary response control sequence that take into account the boundary integrity and the key area blocking accuracy.

[0042] In one example, Dijkstra algorithm is run in the three-dimensional scene graph to calculate the graph distance of each space node to the target path, to obtain the device activation probability value, including: Trajectory prediction is performed on the target current motion state, to obtain a predicted motion trajectory; Each position point in the predicted motion trajectory is matched and positioned in the three-dimensional scene graph to obtain a key space node sequence; Dijkstra algorithm is used to perform traversal calculation with the key space node sequence as the starting point set and using the node topology connection relationship, to obtain the graph distance of each monitoring device node in the scene to the predicted path; The graph distance is converted using an exponential decay function to obtain the device activation probability value.

[0043] In this example, the current motion state of the target is modeled and predicted, considering the current speed and acceleration parameters of the target, and combining historical motion patterns and environmental constraints to deduce the future position of the target. Using a long short-term memory network, a social pooling network, or a trajectory prediction algorithm based on physical constraints, the target's behavior habits and interaction features with the surrounding environment are extracted through time series modeling to obtain a continuous predicted motion trajectory. The predicted trajectory is composed of a series of discrete position points, representing the possible spatial path of the target in the future. Each position point in the predicted trajectory is mapped to the three-dimensional scene graph, and through coordinate matching and range detection, the corresponding spatial nodes are identified. If a position point is located in the intersection area of multiple nodes, it is preferentially assigned to the node that best matches the path connectivity according to the proximity and topological relationship, obtaining a sequence of key spatial nodes, which is equivalent to the spatial node path that the target may pass through in the future. The topological connection relationship of the three-dimensional scene graph is used to calculate the path traversal of the entire graph structure, taking the key nodes on the predicted path as the starting set, and using the Dijkstra algorithm to calculate the shortest graph distance from each monitoring device node to the path. In the execution process of the Dijkstra algorithm, the key nodes are taken as the source points, and all nodes in the scene graph are regarded as target nodes. The shortest reachable distance is calculated based on the edge weight, which can be the physical distance or a weighted index considering the passage weight, floor switching cost, and passage congestion. After the algorithm completes the traversal, the shortest distance value of each monitoring device node relative to the predicted path is output. The graph distance can quantitatively reflect the spatial proximity and monitoring coverage potential between the monitoring device and the target's future trajectory. The graph distance is converted into a device activation probability value, and an exponential decay function is used for probability conversion. When the graph distance between a device node and the predicted path is small, the activation probability should be high, and as the distance increases, the probability decays rapidly. Nonlinear mapping ensures that devices close to the predicted trajectory have a high activation priority, while devices far from the trajectory only retain a low standby probability. Through the above steps, the device activation probability distribution is dynamically generated.

[0044] Referring to Figure 2 The embodiment provides a security monitoring alarm linkage processing system, which comprises: A construction module 1 is configured to construct a three-dimensional scene graph of a monitoring area. A filtering fusion module 2 is configured to acquire visual position information and wireless positioning information of a target, and perform constraint Kalman filtering fusion to obtain a target position. A threat assessment module 3 is configured to map the target position to the three-dimensional scene graph to perform threat assessment and obtain an intelligent alarm signal. A linkage control module 4 is configured to perform linkage control in the three-dimensional scene graph based on the intelligent alarm signal to obtain a device activation instruction and a blocking strategy.

[0045] In the present embodiment, the specific implementation of each unit in the above system embodiment is described in the above method embodiment, which will not be repeated here.

[0046] It should be noted that in this paper, the term "includes", "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, system, article or method including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, system, article or method. Without more limitations, the element defined by the statement "includes a" does not exclude the existence of other identical elements in the process, system, article or method including the element.

[0047] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A security monitoring alarm linkage processing method, characterized in that, The method comprises the following steps: constructing a three-dimensional scene graph of a monitoring area; obtaining visual position information and wireless positioning information of a target, and performing constraint Kalman filtering fusion to obtain a target position; mapping the target position into the three-dimensional scene graph to perform threat assessment and obtain an intelligent alarm signal; performing linkage control in the three-dimensional scene graph based on the intelligent alarm signal to obtain a device activation instruction and a blockade strategy.

2. The security monitoring alarm linkage processing method according to claim 1, characterized in that, The method of constructing a three-dimensional scene graph of a monitoring area comprises the following steps: spatially dividing the monitoring area to obtain a set of spatial nodes; calculating a dynamic safety level value of each spatial node according to the distance of each spatial node in the set of spatial nodes to a key area, a time period danger coefficient and a historical risk score; performing connectivity analysis based on the spatial position relationship of the set of spatial nodes to obtain a node topology connection relationship; performing hierarchical semantic organization on the set of spatial nodes, the node topology connection relationship and the dynamic safety level value to obtain a three-dimensional scene graph.

3. The security monitoring alarm linkage processing method according to claim 1, characterized in that, The method of obtaining visual position information and wireless positioning information of a target, and performing constraint Kalman filtering fusion to obtain a target position comprises the following steps: inputting camera collected images into a YOLO-SecNet neural network to perform multi-task learning detection and obtain camera target detection data; performing pixel coordinate to world coordinate conversion on the boundary box coordinates in the camera target detection data to obtain visual position information of the target in the monitoring scene; performing time difference arrival measurement on the target signal through an ultra-wideband base station to obtain ultra-wideband ranging data; performing non-line-of-sight error compensation and three-dimensional positioning solution on the ultra-wideband ranging data to obtain wireless positioning information; performing constraint Kalman filtering fusion on the visual position information and the wireless positioning information to obtain a target position.

4. The security monitoring alarm linkage processing method according to claim 3, characterized in that, The method of performing constraint Kalman filtering fusion on the visual position information and the wireless positioning information to obtain a target position comprises the following steps: setting the visual position information and the wireless positioning information as camera observation and ultra-wideband observation respectively and constructing a state vector; performing a prediction step and an update step of Kalman filtering based on the state vector to obtain a filtering position result; applying physical constraint conditions of speed limitation and height boundary to the filtering position result to obtain a feasible domain constraint range; performing sequence quadratic programming optimization solution on the filtering position result within the feasible domain constraint range to obtain a target position.

5. The security monitoring alarm linkage processing method according to claim 4, characterized in that, The method of performing sequence quadratic programming optimization solution on the filtering position result within the feasible domain constraint range to obtain a target position comprises the following steps: constructing a constraint minimization problem according to the filtering position result and a precision weighted matrix; combining an objective function of the constraint minimization problem with the feasible domain constraint range to construct an augmented optimization function; performing gradient and Hessian matrix derivation on the augmented optimization function to obtain a search direction vector; determining an optimal iteration step length of the search direction vector through an Armijo line search criterion to obtain a target position.

6. The security monitoring alarm linkage processing method according to claim 1, characterized in that, The method of mapping the target position into the three-dimensional scene graph to perform threat assessment and obtain an intelligent alarm signal comprises the following steps: constructing a trajectory feature vector according to the target position and historical motion data; based on the target position matching corresponding space nodes and extracting the dynamic security level values of the neighborhood nodes in the three-dimensional scene graph, to obtain a scene context feature; inputting the trajectory feature vector and the scene context feature into a threat assessment neural network for spatio-temporal feature fusion, to obtain a probability distribution result; performing numerical calculation on the probability distribution result according to threat level weights, to obtain a target threat level; performing adaptive threshold value determination according to the target threat level, to obtain an intelligent alarm signal.

7. The security monitoring alarm linkage processing method according to claim 6, characterized in that, The adaptive threshold value determination according to the target threat level to obtain an intelligent alarm signal comprises: weighting and fusing the target threat level, the dynamic security level values of the corresponding space nodes and the trajectory behavior features, to obtain an alarm probability; performing likelihood probability calculation on the target motion trajectory by a hidden Markov model to obtain an abnormality index; calculating a reward signal according to statistical data of false alarm events, missed alarm events and accurate alarm events; inputting the reward signal, the alarm probability and the abnormality index into a deep Q network for threshold value decision learning, to obtain an intelligent alarm signal.

8. The security monitoring alarm linkage processing method according to claim 1, characterized in that, The linkage control in the three-dimensional scene graph based on the intelligent alarm signal comprises: performing PID tracking control based on pixel error on a PTZ camera in response to the intelligent alarm signal being triggered, to obtain a first-level response instruction for target locking; running a Dijkstra algorithm in the three-dimensional scene graph to calculate graph distances of each space node to the target path, to obtain a device activation probability value; screening monitoring devices that need to be activated and identifying key passage nodes to perform blocking operations according to the device activation probability value and a preset threshold condition, to obtain a second-level response control sequence and a third-level response control sequence; performing priority sorting on the first-level response instruction, the second-level response control sequence and the third-level response control sequence, to obtain a device activation instruction and a blocking strategy.

9. The security monitoring alarm linkage processing method according to claim 8, characterized in that, The running of the Dijkstra algorithm in the three-dimensional scene graph to calculate graph distances of each space node to the target path, to obtain a device activation probability value, comprises: performing trajectory prediction on a target current motion state, to obtain a predicted motion trajectory; performing space node matching positioning of each position point in the predicted motion trajectory in the three-dimensional scene graph, to obtain a key space node sequence; performing Dijkstra algorithm traversal calculation using node topology connection relationships with the key space node sequence as a starting point set, to obtain graph distances of each monitoring device node to the predicted path in the scene; performing probability conversion based on the graph distances using an exponential decay function, to obtain a device activation probability value.

10. A security monitoring alarm linkage processing system, characterized by, The security monitoring alarm linkage processing system for implementing the steps of the security monitoring alarm linkage processing method of any one of claims 1 to 9 comprises: a construction module for constructing a three-dimensional scene graph of a monitoring area; a filtering fusion module for obtaining visual position information and wireless positioning information of a target and performing constraint Kalman filtering fusion to obtain a target position; The threat assessment module is configured to map the target position to the three-dimensional scene graph for threat assessment, and obtain an intelligent alarm signal. The linkage control module is configured to perform linkage control in the three-dimensional scene graph based on the intelligent alarm signal, and obtain a device activation instruction and a blockade strategy.

Images