Urban rail transit security inspection system

Abstract

The application discloses a kind of urban rail transit security systems, it includes: passenger flow dynamics prediction module is based on crowd pressure distribution and motion vector field constructs correction navier-stokes equation passenger flow model, outputs congestion entropy and risk particle flow;Adaptive millimeter wave scanning array according to risk particle flow intensity and congestion entropy double threshold switch mobile focusing scanning mode or wide area screening mode, the latter generates basic threat index and marks target to be rechecked;Characteristic analysis module generates real-time threat level by contour feature comparison in mobile focusing mode, and threat assessment report is generated by starting lightweight analysis to target to be rechecked in wide area mode;Immune decision center fuses multi-source data, and generates decision antibody dynamically by clone selection algorithm, and the channel topology instruction contained therein is executed according to rules.The application realizes the collaborative optimization of security efficiency and safety in large passenger flow scene by dynamic scanning decision and intelligent channel regulation.

Description

Technical Field

[0001] This invention relates to the field of urban rail transit safety technology. More specifically, this invention relates to an urban rail transit security inspection system. Background Technology

[0002] Urban rail transit security check systems are specialized facilities used at subway and light rail stations to conduct security checks on passengers and their belongings, aiming to prevent prohibited items from entering the transportation system. Current security check systems face a structural contradiction in high-passenger-flow scenarios, where efficiency and security are difficult to optimize synergistically. Existing security check systems generally employ fixed scanning modes, failing to dynamically adjust detection strategies based on real-time passenger flow. This leads to increased congestion during peak hours due to high-precision scanning, or a significantly higher false negative rate under rapid screening modes. Furthermore, static channel designs lack physical responsiveness to sudden risks, and reliance on manual identification and guidance results in response delays, making it difficult to isolate high-risk targets in a timely manner. Insufficient accuracy of passenger flow prediction models is a core constraint. Traditional methods struggle to construct effective real-time assessment systems, primarily due to the difficulty in quantifying and modeling the strong nonlinear characteristics of crowd flow, and the inefficient fusion processing of multi-source heterogeneous data (such as pressure distribution and motion vectors). The lack of coordination mechanisms among the functional modules of existing security inspection systems leads to a disconnect between decision-making and execution. Channel resource allocation cannot adapt to fluctuations in passenger flow, especially under sudden surges in passenger volume, which can easily cause system chaos. Attempts to introduce automated control fail to balance passage efficiency and threat interception accuracy due to a lack of objective indicators for quantifying risk propagation and adaptive channel topology technology. Specifically, scanning mode switching relies on manual experience and lacks a quantitative triggering mechanism based on risk propagation characteristics; threat handling procedures are rigid, failing to implement tiered physical isolation based on risk levels; and the channel structure is static and fixed, unable to dynamically optimize topology in response to changes in congestion entropy. The aforementioned defects together have forced rail transit security inspection systems to sacrifice security inspection accuracy or passage efficiency under the pressure of large passenger flows, forming a long-standing unresolved technical bottleneck. Summary of the Invention

[0003] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0004] Another objective of this invention is to provide an urban rail transit security inspection system that solves the technical problem of balancing security inspection efficiency and safety in scenarios with large passenger flows.

[0005] To achieve these objectives and other advantages according to the present invention, an urban rail transit security inspection system is provided, comprising: The passenger flow dynamics prediction module, based on the collected passenger flow pressure distribution and motion vector field, constructs a passenger flow micro-environment model based on the modified Navier-Stokes equations, and outputs congestion entropy values ​​and risk particle flow in real time; An adaptive millimeter-wave scanning array receives congestion entropy values ​​and risk particle streams. When the intensity of the risk particle stream is greater than a predetermined intensity threshold, a moving focusing scanning mode is activated. When the intensity of the risk particle stream is less than or equal to the predetermined intensity threshold and the congestion entropy value is greater than the predetermined entropy threshold, a wide-area screening mode is activated, and a basic threat index is generated. When a passenger's basic threat index is found to be greater than a safety threshold, the passenger is marked as a target to be re-examined; otherwise, the wide-area screening mode is executed. The feature analysis module, in the mobile focus scanning mode, receives and parses millimeter-wave scanning data, extracts the outline features of the items carried by the focused passenger, compares the outline features with the pre-stored contraband feature database, and generates the real-time threat level of the passenger; in the wide-area screening mode, when the target identification to be re-inspected is received, the lightweight analysis model is started to analyze the millimeter-wave data of the target to be re-inspected and generate a threat assessment report of the target to be re-inspected. The immune decision center receives congestion entropy values, real-time threat levels, and threat assessment reports for targets awaiting re-inspection. It dynamically generates decision antibodies using a cloning selection algorithm. These decision antibodies include channel topology instructions and a set of device parameters. The generation rules for channel topology instructions include: if the threat information is at a high-risk real-time threat level, an instruction containing a 30° acute-angle diversion is forcibly generated; if the threat information is at a non-high-risk real-time threat level, a corresponding channel topology instruction is generated based on the predetermined entropy range where the congestion entropy value is located; when a target identifier and its corresponding threat assessment report are received, a channel topology instruction pointing to the secondary diversion area is generated.

[0006] Preferably, the urban rail transit security inspection system further includes: The variable topology channel group receives and executes channel topology instructions and equipment parameter sets, dynamically adjusting the channel physical structure; based on instructions including 30° acute-angle flow dividers, it forms acute-angle flow divider structures at corresponding positions; based on instructions pointing to secondary flow dividers, it guides the target to be re-inspected to its preset secondary flow divider area; within the secondary flow divider area, it further constructs the channel structure based on channel topology instructions generated from the threat assessment report for the target to be re-inspected; when the threat assessment report indicates high risk, it constructs a π-shaped aggregate channel to guide the target to be re-inspected to the re-inspection area; it collects real-time data on changes in pedestrian flow pressure within the channel and feeds it back to the passenger flow dynamics prediction module for calibrating the passenger flow microenvironment dynamics model.

[0007] Preferably, in the urban rail transit security inspection system, the intensity value of the risk particle stream is dynamically generated through the following steps: S1.1 Real-time access to the risk path spatial distribution data output by the passenger flow dynamics prediction module, synchronously acquiring the mobile device signaling density distribution data provided by the mobile communication base station, and the gesture skeleton key point vector direction captured by the security inspector's wearable device; S1.2. Based on mobile device signaling density distribution data, identify the boundary coordinates of abnormal clustering areas; based on risk path spatial distribution data, extract the path vector of the main diffusion direction; S1.3 Calculate the percentage of overlap between the boundary coordinates of the abnormal clustering area and the spatial distribution area of ​​the risk path, and use it as the first correction coefficient; calculate the cosine of the angle between the direction of the gesture skeleton key point vector and the direction of the path vector of the main diffusion direction, and use it as the second correction coefficient. S1.4. The frequency of occurrence of historical prohibited items stored in the security check event database is used as the basic weight. Combined with the first correction coefficient and the second correction coefficient, it is input into the pre-trained risk propagation tree model. The risk propagation tree model is generated by pre-training based on the spatiotemporal distribution pattern of the historical security check event database. S1.5 The risk propagation tree model outputs the updated risk particle flow intensity value and feeds it back to the passenger flow dynamics prediction module.

[0008] Preferably, the urban rail transit security inspection system further includes: The dynamic path calibration engine connects to the initial path output of the risk particle flow of the passenger flow dynamics prediction module and is connected in parallel to the multispectral interference sensing array deployed on the tunnel arch. The multispectral interference sensing array simultaneously collects the physiological agitation index in the thermal infrared band and the motion change trajectory in the visible light band. When the physiological agitation index exceeds the preset agitation threshold and lasts for more than 5 seconds, or when the cumulative deviation of the azimuth angle between the sudden movement trajectory and the initial path is greater than or equal to 30°, the nonlinear compensator is activated. Once the nonlinear compensator is activated, it performs the following operations: S3.1 Map the physiological agitation index to a path offset probability density function, and generate the first compensation vector through a convolution kernel; S3.2 Decompose the tangential and normal acceleration components of the trajectory of sudden motion change, and generate a second compensation vector with the amplitude of the normal component as the weight; S3.3, merge the first compensation vector and the second compensation vector into the initial path of the risk particle flow, and output the anti-disturbance calibration path to the risk path input of the adaptive millimeter-wave scanning array.

[0009] Preferably, the urban rail transit security inspection system continuously monitors antibody activity after the immune decision-making center generates decision antibodies. S4.1. Real-time acquisition of threat level update flags and instruction execution progress values ​​of the topology variable channel group from the feature analysis module; S4.2 When the progress value of the channel group executing the high-risk acute angle fluid splitting command is less than 50%, if the threat level update flag is switched to the medium-low risk state, the antibody apoptosis protocol is activated. S4.3 The antibody apoptosis protocol uses the instruction execution progress value as the decay base and combines it with the updated risk particle flux intensity value to generate the antibody half-life; S4.4 When the instruction execution progress value exceeds the progress threshold corresponding to the antibody half-life, the clone selection algorithm is forcibly triggered to recombinate the decision antibody. The recombination rule uses non-acute angle diversion instructions generated by the entropy value interval to cover the original high-risk instructions.

[0010] Preferably, the urban rail transit security inspection system further includes a real-time video compensation unit, which continuously collects passenger flow video streams through a wide-angle camera array deployed on the side wall of the passage. The real-time video compensation unit performs the following operations: S5.1 Extract the passenger movement speed distribution matrix and direction change rate in the video stream using optical flow method; S5.2 When the standard deviation of the moving speed exceeds the preset speed tolerance and the rate of change of direction is greater than the direction threshold, an emergency topology pre-instruction is generated; S5.3 Input emergency topology pre-instructions in parallel to the immune decision center and the topology variable channel group; Among them, when the passenger flow dynamics prediction module outputs a delay, the immune decision center uses the emergency topology pre-instruction as an antigen input substitute to drive the clonal selection algorithm to generate decision antibodies containing channel topology instructions; after receiving the emergency topology pre-instruction, the topology variable channel group generates channel topology instructions according to the entropy value interval.

[0011] Preferably, in the urban rail transit security inspection system, the real-time video compensation unit further performs the following operations: S5.4 Continuously monitor the acquisition status of mobile communication base station signaling density distribution data and gesture skeleton key point vectors; S5.5 When the duration of missing signaling density distribution data exceeds the first tolerance or the invalid count of gesture skeletal keypoint vectors exceeds the second tolerance, activate the video compensation enhancement protocol and perform the following operations: The video stream collected by the wide-angle camera array is used to generate a passenger flow concentration distribution map through spatiotemporal slicing analysis, which replaces the missing signaling density distribution data; The passenger's upper limb movement trajectory is analyzed using a pose estimation algorithm to generate a simulated gesture skeletal key point vector; S5.7 Input the passenger flow concentration distribution map and the simulated gesture skeleton key point vector into the risk particle flow intensity generation step.

[0012] Preferably, the urban rail transit security inspection system further includes a multimodal data redundancy engine, the input of which is connected to the real-time signaling data interface of the existing surveillance camera group and mobile communication base station in the station. The multimodal data redundancy engine performs the following operations: S6.1 Continuously monitor the equipment status codes of the ground pressure sensing network and 3D lidar. When the status code remains abnormal for a duration longer than the fault tolerance limit, activate the data reconstruction protocol. S6.2 After the data reconstruction protocol is activated, perform the following operations: a) Analyze the video streams from the surveillance camera group and generate a heat map of passenger flow density distribution using the background subtraction method; b) Extract the signaling hopping trajectory of the mobile communication base station and construct the passenger movement direction probability matrix; S6.3 Map the passenger flow density distribution heat map to simulated pressure distribution data, convert the passenger movement direction probability matrix into simulated motion vector field, and input it into the passenger flow dynamics prediction module to replace the original data source; S6.4 When the device status code returns to normal, it will automatically switch back to the original data acquisition mode.

[0013] Preferably, in the urban rail transit security inspection system, when simultaneously acquiring mobile communication base station signaling density distribution data and gesture skeleton key point vector direction in step S1.1, if the effective coverage rate of the signaling density distribution data is less than a preset coverage threshold or the confidence level of the gesture skeleton key point vector is less than a confidence threshold, the multi-source biometric fusion protocol is activated. The multi-source biometric fusion protocol performs the following operations: The phase difference of passenger gait cycles is collected by an infrared thermal imaging array deployed at the top of the passage, generating a gait aggregation density distribution map, which is then normalized to replace the signaling density distribution data. The direction of passenger torso swing is analyzed by the micro-Doppler effect of millimeter-wave scanning array, and a biomechanical direction vector is generated. When the swing amplitude is greater than the sensitivity threshold, the gesture skeletal key point vector is replaced. The gait aggregation density distribution map is input into the abnormal aggregation region identification process in step S1.2, and the biomechanical direction vector is input into the second correction coefficient calculation process in step S1.3.

[0014] The present invention has at least the following beneficial effects: 1. The urban rail transit security inspection system provided by this invention constructs a passenger flow micro-environment model based on the modified Navier-Stokes equations through a passenger flow dynamics prediction module, and outputs quantitative congestion entropy values ​​and risk particle flows in real time, solving the defect of traditional passenger flow prediction models that are difficult to accurately model due to nonlinear characteristics; the adaptive millimeter-wave scanning array dynamically switches between moving focusing scanning and wide-area screening modes based on dual thresholds, realizing the precise deployment of scanning resources according to risk levels, significantly reducing the risk of missed detection while ensuring basic passage efficiency for large passenger flows; the immune decision center uses a clonal selection algorithm to generate decision antibodies, and forcibly triggers 30° acute angle flow splitting commands for high-risk real-time threat levels, it achieves second-level isolation and blocking of high-risk targets through physical topology, overcoming the control loopholes caused by the lag in manual response; combined with the dynamic path calibration engine for real-time anti-interference calibration of risk particle flow paths, and the antibody activity monitoring mechanism for dynamic correction of the command execution process, a "prediction-decision-execution-feedback" closed-loop system is formed, ultimately breaking through the technical bottleneck of the long-term difficulty in coordinating the optimization of security inspection efficiency and safety in large passenger flow scenarios; 2. This invention utilizes a passenger flow dynamics prediction module to output congestion entropy and risk particle flow in real time based on the modified Navier-Stokes equations, providing a quantitative basis for scanning decisions. An adaptive millimeter-wave scanning array dynamically switches between moving-focus scanning mode and wide-area screening mode based on dual thresholds of risk particle flow intensity and congestion entropy, enabling precise allocation of scanning resources according to risk levels. A feature analysis module outputs real-time threat levels and threat assessment reports in a tiered manner. An immune decision-making center generates differentiated channel topology instructions based on a clonal selection algorithm, solving the efficiency and safety imbalance problem caused by fixed scanning strategies and static channels in traditional systems under high passenger flow conditions.

[0015] 3. The variable topology channel group of the present invention executes the 30° acute angle flow diversion command triggered by high-risk targets, forming a physical isolation structure to block the movement path of high-risk targets; for targets to be re-inspected, a π-shaped aggregation channel is constructed in the secondary diversion area to guide them to the re-inspection area, realizing the gradient control of threatening targets; by real-time feedback of people flow pressure data to calibrate the prediction model, a closed-loop optimization mechanism is formed to improve the system's adaptive capability.

[0016] 4. This invention identifies the boundaries of abnormal clustering areas by using mobile signaling data, calculates dual correction coefficients by combining the main diffusion direction of the risk path with the skeletal vector of the security inspector's gesture, and dynamically generates risk particle flow intensity values ​​using a risk propagation tree model. This solves the problem of insufficient accuracy in traditional risk prediction that relies on a single data source, and improves the accuracy of high-risk area identification.

[0017] 5. The dynamic path calibration engine of this invention activates a nonlinear compensator based on the thermal infrared physiological agitation index and visible light motion mutation trajectory. By generating vectors to compensate for the initial path of risky particle flow, it overcomes the path deviation defects caused by environmental interference and ensures the positioning accuracy of the adaptive millimeter-wave scanning array.

[0018] 6. When the antibody activity monitoring mechanism of this invention has not reached 50% of the instruction execution progress, it dynamically triggers the antibody apoptosis protocol according to the threat level downgrade, generates the antibody half-life with the progress value as the decay base, and recombines the non-acute angle diversion instruction to cover the original high-risk instruction, thus avoiding the waste of channel resources caused by excessive control.

[0019] 7. The real-time video compensation unit of the present invention extracts the characteristics of sudden changes in passenger flow speed and direction using optical flow method to generate emergency topology pre-instructions. When the passenger flow dynamics prediction module is delayed, it drives the immune decision center and channel group to respond to emergencies, preventing the system from failing to make decisions due to data delay.

[0020] 8. When signaling or gesture data fails, the video compensation enhancement protocol of the present invention generates a passenger flow aggregation distribution map through spatiotemporal slicing analysis, and combines it with posture estimation to simulate gesture skeletal vectors, thereby ensuring the data integrity of the risk particle flow intensity calculation process and enhancing the system's fault tolerance.

[0021] 9. When the pressure sensing network or lidar fails, the multimodal data redundancy engine of the present invention uses monitoring video to generate a passenger flow density heat map and signal jump trajectory to construct a simulated motion vector field, maintain the data input stability of the passenger flow dynamics prediction module, and ensure the continuous operation of the system.

[0022] 10. The multi-source biometric fusion protocol of the present invention generates a normalized gait aggregation density map to replace signaling data by using infrared thermal imaging gait phase difference, and extracts the torso swing direction vector (when the swing amplitude meets the standard) based on millimeter wave micro-Doppler effect to replace gesture data, thereby solving the data reliability problem when the biometric collection environment is limited.

[0023] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to embodiments, so that those skilled in the art can implement it based on the description.

[0025] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0026] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0027] This invention provides a security inspection system for urban rail transit, comprising: The passenger flow dynamics prediction module constructs a passenger flow micro-environment model based on the collected passenger flow pressure distribution and motion vector field, and outputs congestion entropy value and risk particle flow in real time. The congestion entropy value is a scalar indicator that quantifies the degree of regional congestion, and the risk particle flow is a vector field that characterizes the direction and intensity of risk propagation. The moving focused scanning mode refers to the high-precision scanning of specific passengers by the millimeter-wave array, and the wide-area screening mode refers to the rapid scanning of the crowd to generate the basic threat index. An adaptive millimeter-wave scanning array receives congestion entropy values ​​and risk particle streams. When the intensity of the risk particle stream is greater than a predetermined intensity threshold, a moving focusing scanning mode is activated. When the intensity of the risk particle stream is less than or equal to the predetermined intensity threshold and the congestion entropy value is greater than the predetermined entropy threshold, a wide-area screening mode is activated, and a basic threat index is generated. When a passenger's basic threat index is found to be greater than a safety threshold, the passenger is marked as a target to be re-examined; otherwise, the wide-area screening mode is executed. The feature analysis module, in the mobile focus scanning mode, receives and parses millimeter-wave scanning data, extracts the outline features of the items carried by the focused passenger, compares the outline features with the pre-stored contraband feature database, and generates the real-time threat level of the passenger; in the wide-area screening mode, when the target identification to be re-inspected is received, the lightweight analysis model is started to analyze the millimeter-wave data of the target to be re-inspected and generate a threat assessment report of the target to be re-inspected. The immune decision center receives congestion entropy values, real-time threat levels, and threat assessment reports for targets awaiting re-inspection. It dynamically generates decision antibodies using a cloning selection algorithm. These decision antibodies include channel topology instructions and a set of device parameters. The generation rules for channel topology instructions include: if the threat information is at a high-risk real-time threat level, an instruction containing a 30° acute-angle diversion is forcibly generated; if the threat information is at a non-high-risk real-time threat level, a corresponding channel topology instruction is generated based on the predetermined entropy range where the congestion entropy value is located; when a target identifier and its corresponding threat assessment report are received, a channel topology instruction pointing to the secondary diversion area is generated.

[0028] In the above technical solution, the passenger flow dynamics prediction module of the urban rail transit security inspection system provided by the present invention continuously collects the passenger flow pressure distribution and motion vector data in the channel through ground pressure sensors and three-dimensional lidar. By modifying the Navier-Stokes equations, it simulates the fluid dynamics behavior of passenger flow and outputs in real time the congestion entropy value that quantifies the degree of regional congestion and the risk particle flow that marks the risk propagation path.

[0029] The adaptive millimeter-wave scanning array makes dynamic decisions based on the received congestion entropy value and risk particle flow parameters: when the intensity of the risk particle flow exceeds a predetermined intensity threshold, it immediately locks onto the high-risk propagation direction and initiates a moving focusing scanning mode, driving a rotating gimbal to lock onto high-risk targets for precise millimeter-wave scanning; if the intensity of the risk particle flow is equal to or lower than the predetermined intensity threshold (not exceeding the limit), and the congestion entropy value exceeds the entropy threshold, it switches to a wide-area screening mode, quickly scanning the crowd through a wide-beam transmitter to generate a basic threat index, and marking passengers with an index exceeding the limit as targets to be re-examined; when both the congestion entropy value and the risk particle flow parameters are within the limits, it maintains basic wide-area screening to ensure traffic efficiency.

[0030] The feature analysis module parses the scan data (millimeter-wave scan data) in moving focus mode and generates a real-time threat level by comparing the contour features with a pre-stored contraband database; in wide-area mode, it starts a lightweight model for the target to be re-inspected and outputs a threat assessment report.

[0031] The immune decision-making center integrates congestion entropy values, real-time threat levels, and threat assessment reports, and uses a clonal selection algorithm to generate dynamic decision antibodies containing channel topology instructions and equipment parameter sets. Channel topology instructions are generated according to rules: high-risk real-time threat levels forcibly trigger isolation instructions containing 30° acute-angle diversions; non-high-risk threats adapt the channel layout based on the congestion entropy value's location; targets awaiting re-inspection uniformly trigger guidance instructions pointing to secondary diversion areas.

[0032] Traditional security screening systems, due to their fixed scanning patterns and static channel designs, face a trade-off between efficiency and security under high passenger flow conditions—high-precision scanning leads to congestion, rapid screening increases the risk of missed detections, and reliance on manual response results in delays. This solution uses passenger flow modeling to predict risk areas and dynamically allocate scanning resources; a cloning selection algorithm generates topology commands for second-level threat response; and a 30° acute-angle flow divider physically blocks high-risk targets. Entropy range adaptation to channel layout continuously optimizes throughput efficiency, and combined with pressure data from channel groups, the predictive model is calibrated in real time to form a closed-loop optimization, ultimately achieving a synergistic improvement in security screening efficiency and security under high passenger flow scenarios.

[0033] The above technical solution achieves dynamic adaptation between scanning strategies and passenger flow status: It precisely switches scanning modes through a dual-threshold trigger mechanism, quickly locking high-risk targets when risk particle flow exceeds limits, initiating wide-area screening to prevent channel congestion when congestion entropy values ​​rise, and maintaining basic throughput efficiency when both parameters are normal. A hierarchical analysis mechanism ensures a balance between in-depth detection of high-threat targets and rapid classification of targets to be inspected. Differentiated channel instructions generated by the clone selection algorithm achieve three levels of control: physical isolation of high-risk targets, entropy-based diversion of non-high-risk targets, and directional guidance of targets to be inspected. This comprehensively resolves the core contradiction of the inability to simultaneously achieve both high false negative rates and throughput efficiency under a fixed scanning strategy.

[0034] In another technical solution, the urban rail transit security inspection system further includes: The variable topology channel group receives and executes channel topology instructions and equipment parameter sets, dynamically adjusting the channel physical structure; based on instructions including 30° acute-angle flow dividers, it forms acute-angle flow divider structures at corresponding positions; based on instructions pointing to secondary flow dividers, it guides the target to be re-inspected to its preset secondary flow divider area; within the secondary flow divider area, it further constructs the channel structure based on channel topology instructions generated from the threat assessment report for the target to be re-inspected; when the threat assessment report indicates high risk, it constructs a π-shaped aggregate channel to guide the target to be re-inspected to the re-inspection area; it collects real-time data on changes in pedestrian flow pressure within the channel and feeds it back to the passenger flow dynamics prediction module for calibrating the passenger flow microenvironment dynamics model.

[0035] Traditional urban rail transit security check systems employ fixed physical lane designs, making it impossible to dynamically adjust the layout based on risk levels. When high-risk individuals are detected, security personnel manually guide them to a re-inspection area, leading to blind spots in control due to delays in manual response. Furthermore, individuals awaiting inspection share lanes with regular passengers, causing cross-congestion and efficiency losses. More critically, the allocation of lane resources lacks real-time passenger flow data support, making it impossible to dynamically optimize in response to changes in congestion entropy, resulting in a break in the decision-making and execution chain. For example, static lanes are forced to extend queues during peak hours, while lane resources remain idle during off-peak hours, making it difficult to improve resource utilization and traffic efficiency in tandem.

[0036] This solution achieves dynamic physical layer response through a topologically variable channel group. When a high-risk command is received, the hydraulic drive mechanism pushes a baffle to form a 30° acute-angle flow-diverting structure, creating a physical isolation zone in front of the target passenger's path, forcibly preventing contact with other passengers. After the baffle closes, it automatically locks until the re-inspection is completed. Targets awaiting inspection are guided to an internal secondary diversion area, where they are handled according to a tiered approach based on the threat assessment report—high-risk targets trigger a three-sided movable barrier to construct a π-shaped closed channel, achieving point-to-point closed transfer; non-high-risk targets automatically adjust the channel width according to the real-time congestion entropy value. For example, at high entropy values, the channel narrows to accelerate diversion, while at low entropy values, the width expands to reduce density. Embedded pressure sensors within the channels collect real-time changes in passenger flow pressure, mapping this to a pressure gradient matrix that is fed back to the prediction module. This continuously calibrates the viscosity coefficient parameters in the passenger flow model, forming a closed-loop enhancement between execution effectiveness and model optimization.

[0037] Compared to existing technologies, this solution achieves second-level physical isolation of high-risk targets through a dynamic channel structure, increasing response speed several times faster than manual guidance and completely eliminating blind spots in control. The π-shaped enclosed channel prevents the intersection of targets awaiting inspection with normal passenger flow, reducing the risk of secondary inspections. Dynamic adjustment of channel width based on entropy values ​​ensures precise matching of capacity to passenger load, improving peak-hour turnaround efficiency. A pressure feedback-driven closed-loop optimization mechanism continuously improves the accuracy of congestion entropy calculation over time; pilot data from a subway station shows a 40% reduction in prediction error rate after 30 days. These substantial improvements, for the first time, unify risk control and resource scheduling at the physical execution layer, providing reliable technical support for the security inspection system.

[0038] This technical solution constructs a closed-loop control system at the physical execution layer. A 30° acute-angle flow divider uses a rigid baffle structure to achieve hard isolation of high-risk targets, blocking their contact with other passenger paths. A π-shaped aggregation channel forms a closed transfer channel for high-risk targets to be inspected, avoiding secondary risks during re-inspection. The channel width within the secondary diversion zone is dynamically adjusted according to entropy values, optimizing traffic efficiency in real time. Pressure feedback data continuously calibrates the prediction model parameters, ensuring that the calculation accuracy of congestion entropy and risk particle flow continuously improves as the system operates, forming a self-reinforcing mechanism of "command execution - environmental feedback - model optimization".

[0039] The specific operating steps of the security inspection system for rail transit provided by this invention are as follows: Step 1: Passenger Flow Dynamic Forecast Real-time data collection of pedestrian flow pressure distribution and motion vector data within the passageway is used to construct a passenger flow fluid model by modifying the Navier-Stokes equations. Two key parameters are calculated and output: congestion entropy value - quantifying the degree of regional congestion; and risk particle flow - marking the direction and intensity of risk propagation.

[0040] Step Two: Dynamic Scanning Decision Switch the scan mode according to the parameters output in step one: If the intensity of the risk particle flow is greater than the predetermined intensity threshold: activate the moving focus scanning mode and perform millimeter-wave depth scanning on the high-risk passengers; If the intensity of the risk particle flow is less than or equal to the intensity threshold and the congestion entropy value is greater than the entropy threshold: activate the wide-area screening mode, quickly scan the crowd to generate a basic threat index, and mark passengers whose index is greater than the safety threshold as targets to be re-examined. If both parameters are less than or equal to the threshold: maintain the basic wide-area screening mode.

[0041] Step 3: Threat Classification Analysis In mobile focus mode: analyze the millimeter-wave data of the locked passenger, extract the outline features of the items, compare them with the contraband database, and generate a real-time threat level (high risk / medium risk / low risk). In wide-area screening mode: a lightweight model is activated for targets to be re-examined, and a threat assessment report is quickly generated.

[0042] Step 4: Intelligent Channel Decision Generate decision instructions using the clonal selection algorithm: If the real-time threat level is high: force the generation of a channel command containing a 30° acute angle fluid splitter; If the real-time threat level is not high-risk: Generate an appropriate channel instruction based on the range of congestion entropy value; If there is a target to be re-inspected: generate a boot command pointing to the secondary shunt area.

[0043] Step 5: Physical Channel Execution High-risk command response: Create a 30° acute-angle flow-diverting structure at the designated location to physically block the target path; Response to pending re-inspection order: Guide the target to the secondary diversion area and construct an intra-area corridor based on the threat assessment report: The report indicates high risk: a π-shaped aggregation channel is generated and the patient is guided to the re-examination area. Other cases: Adjust the channel layout according to the entropy value command (such as widening / shrinking the path).

[0044] Step Six: Closed-Loop Model Calibration Real-time data on changes in passenger flow pressure after channel structure adjustment is collected and fed back to the passenger flow dynamics prediction module to dynamically calibrate fluid model parameters.

[0045] In another technical solution, the intensity value of the risk particle stream in the urban rail transit security inspection system is dynamically generated through the following steps: S1.1 Real-time access to the risk path spatial distribution data output by the passenger flow dynamics prediction module, synchronously acquiring the mobile device signaling density distribution data provided by the mobile communication base station, and the gesture skeleton key point vector direction captured by the security inspector's wearable device; S1.2. Based on mobile device signaling density distribution data, identify the boundary coordinates of abnormal clustering areas; based on risk path spatial distribution data, extract the path vector of the main diffusion direction; S1.3 Calculate the percentage of overlap between the boundary coordinates of the abnormal clustering area and the spatial distribution area of ​​the risk path, and use it as the first correction coefficient; calculate the cosine of the angle between the direction of the gesture skeleton key point vector and the direction of the path vector of the main diffusion direction, and use it as the second correction coefficient. S1.4. The frequency of occurrence of historical prohibited items stored in the security check event database is used as the basic weight. Combined with the first correction coefficient and the second correction coefficient, it is input into the pre-trained risk propagation tree model. The risk propagation tree model is generated by pre-training based on the spatiotemporal distribution pattern of the historical security check event database. S1.5 The risk propagation tree model outputs the updated risk particle flow intensity value and feeds it back to the passenger flow dynamics prediction module.

[0046] Traditional risk prediction models typically rely on video analytics or Wi-Fi location data to obtain crowd density distribution and trigger alarms based on preset density thresholds. These methods cannot distinguish between normal queuing and abnormal crowding. For example, Beijing West Railway Station once mistakenly triggered 12 high-risk alarms due to excessive queue density during the morning rush hour, causing frequent passageway closures. More importantly, the models do not consider the synergy between risk propagation direction and human intervention; the predicted path often deviates from actual handling needs, forcing security personnel to manually correct system commands, resulting in response delays of up to minutes.

[0047] This solution reconstructs the risk particle flow intensity calculation process through multi-source data fusion: It accesses the risk path spatial distribution data output by the passenger flow prediction module in real time, simultaneously acquires the mobile base station signaling density distribution and the gesture skeleton vectors captured by the security personnel's wearable devices, identifies the boundary coordinates of abnormal clustering areas based on signaling density, and extracts the main diffusion direction vector by combining risk path data. The overlap ratio between the abnormal area boundary and the risk path space is calculated as the first correction coefficient, reflecting the spatial matching degree; the cosine of the angle between the gesture skeleton vector and the main path vector is calculated as the second correction coefficient, quantifying the consistency between human intervention and system prediction. The frequency of historical contraband occurrences is used as the basic weight, and the two correction coefficients are integrated into a pre-trained risk propagation tree model. This model is trained based on the spatiotemporal patterns of hundreds of thousands of historical security check events, and finally outputs dynamically calibrated risk particle flow intensity values, which are then fed back to the prediction module.

[0048] Compared to existing technologies, this solution overcomes the limitations of a single data dimension. It addresses the problem of misclassifying ordinary queues as high-risk areas through a spatial overlap coefficient, and achieves second-level alignment between human intervention intentions and system predictions through a directional cosine coefficient. The risk propagation tree model integrates static historical patterns with dynamic environmental corrections, ensuring that risk intensity values ​​always adapt to real-time scenarios. The intervention dimension introduced by security inspector gesture vectors enables the system to have human-machine collaborative evolution capabilities, reducing the false alarm rate to one-fifth of traditional methods. Ultimately, it ensures the accuracy of high-risk area identification while avoiding the waste of channel resources caused by false alarms.

[0049] The aforementioned technical solution improves the reliability of multi-source fusion in risk prediction. Mobile signaling data identifies macroscopic abnormal clustering areas, risk path data provides the main diffusion direction, and security inspector gesture vectors introduce a human intervention dimension. The three data sources are quantitatively corrected through overlapping area ratio (spatial consistency) and direction cosine (behavioral consistency). Historical contraband frequency weights establish a risk propagation probability benchmark, and the pre-trained model integrates dynamic correction coefficients to output intensity values. This solves the problem of misjudging high-risk areas due to missing data dimensions in traditional single-sensor prediction, significantly reducing the risk location error rate.

[0050] In another technical solution, the urban rail transit security inspection system further includes: The dynamic path calibration engine connects to the initial path output of the risk particle flow of the passenger flow dynamics prediction module and is connected in parallel to the multispectral interference sensing array deployed on the tunnel arch. The multispectral interference sensing array simultaneously collects the physiological agitation index in the thermal infrared band and the motion change trajectory in the visible light band. When the physiological agitation index exceeds the preset agitation threshold and lasts for more than 5 seconds, or when the cumulative deviation of the azimuth angle between the sudden movement trajectory and the initial path is greater than or equal to 30°, the nonlinear compensator is activated. Once the nonlinear compensator is activated, it performs the following operations: S3.1 Map the physiological agitation index to a path offset probability density function, and generate the first compensation vector through a convolution kernel; S3.2 Decompose the tangential and normal acceleration components of the trajectory of sudden motion change, and generate a second compensation vector with the amplitude of the normal component as the weight; S3.3, merge the first compensation vector and the second compensation vector into the initial path of the risk particle flow, and output the anti-disturbance calibration path to the risk path input of the adaptive millimeter-wave scanning array.

[0051] Existing risk path calibration systems typically rely on a single sensor (such as LiDAR) to detect positional deviations and use preset angle thresholds to trigger mechanical corrections. For example, an airport security system failed to detect a passenger's sudden fainting spell causing localized congestion, mistakenly identifying physiological stagnation as the end of the path, resulting in subsequent scans focusing on the wrong area for up to 3 minutes. More critically, traditional linear compensators cannot handle the acceleration vector decomposition in sudden motion changes; during calibration, they simply translate the path line, ignoring the differential effects of tangential and normal acceleration on the offset direction.

[0052] This solution reconstructs the calibration process through multispectral interference sensing and nonlinear compensation. A thermal infrared camera deployed on the tunnel arch continuously collects the rate of temperature rise of human skin, outputting a physiological agitation index; a visible light camera simultaneously captures trajectories of sudden stops, reverse movements, and other abrupt changes in motion. The dynamic path calibration engine activates the nonlinear compensator in two scenarios: when the physiological agitation index exceeds a threshold and lasts for 5 seconds, it reflects the disordered group behavior caused by emotional contagion; when the cumulative deviation of the abrupt motion trajectory from the initial path azimuth angle reaches 30 degrees, it indicates that the behavior pattern has deviated from the prediction framework. The compensator maps the physiological agitation index to a probability density function, generates a first compensation vector through spatial convolution, quantifies the spatial pushing effect of emotional fluctuations on the path; decomposes the tangential and normal acceleration components of the abrupt motion trajectory, and generates a second compensation vector with the normal component amplitude as the weight, capturing the intensity of centrifugal disturbances caused by sudden changes in direction. The two vectors are fused to the initial path of the risk particle flow, and an anti-interference calibration path is output to the millimeter-wave scanning array.

[0053] Compared to traditional methods, this approach overcomes the limitations of rigid and one-dimensional compensation. For the first time, the multispectral synergy of thermal infrared and visible light correlates physiological agitation with path deviation, preventing the misjudgment of emotional agglomeration as risk propagation. The nonlinear compensator precisely quantifies the direction of behavioral mutations through acceleration vector decomposition, ensuring the calibration path closely matches the actual risk diffusion situation. The false locking rate of the millimeter-wave scanning array due to interference is significantly reduced, and the focusing accuracy of high-risk targets reaches the level of human expert judgment, fundamentally curbing systemic resource waste.

[0054] The aforementioned technical solution enhances the anti-interference capability of risky paths. The thermal infrared physiological agitation index (>0.5℃ / s temperature rise) reflects sudden changes in crowd emotions, while visible light motion abrupt change trajectories (sudden stops / reverse movements) capture behavioral anomalies. Both serve as environmental interference indicators to trigger a compensation mechanism. The path offset probability function converts physiological parameters into spatial offsets, and the weighted normal acceleration reflects the intensity of behavioral abrupt changes. Vector fusion compensates for initial path deviations. Ultimately, this ensures that the focusing and positioning accuracy of the millimeter-wave scanning array remains above 90% even in environments with strong interference, avoiding misscanning or missed scans caused by environmental noise.

[0055] In another technical solution, the urban rail transit security inspection system continuously monitors antibody activity after the immune decision-making center generates decision antibodies. S4.1. Real-time acquisition of threat level update flags and instruction execution progress values ​​of the topology variable channel group from the feature analysis module; S4.2 When the progress value of the channel group executing the high-risk acute angle fluid splitting command is less than 50%, if the threat level update flag is switched to the medium-low risk state, the antibody apoptosis protocol is activated. S4.3 The antibody apoptosis protocol uses the instruction execution progress value as the decay base and combines it with the updated risk particle flux intensity value to generate the antibody half-life; S4.4 When the instruction execution progress value exceeds the progress threshold corresponding to the antibody half-life, the clone selection algorithm is forcibly triggered to recombinate the decision antibody. The recombination rule uses non-acute angle diversion instructions generated by the entropy value interval to cover the original high-risk instructions.

[0056] Current security systems employ a one-way trigger mode, executing a high-risk lane command indefinitely once it's generated. A subway station once experienced a misjudgment: a passenger's bag containing metal tableware triggered the high-risk command. After a sharp-angle flow control system initiated a closure and a re-inspection confirmed it was non-threatening, the system couldn't automatically deactivate the command, causing the lane closure to continue until manual intervention. During peak hours, this resulted in a cumulative waste of 15 minutes of traffic resources. This rigid execution mode causes chain congestion in high-traffic scenarios and fails to respond to real-time risk downgrading.

[0057] This solution reconstructs instruction lifecycle management through antibody activity monitoring, acquiring real-time threat level updates from the feature analysis module and instruction execution progress values ​​for the channel group. When the execution progress of a high-risk acute-angle shunt instruction falls below 50% and the threat level is downgraded to low-to-medium risk, an antibody apoptosis protocol is activated. This protocol uses the current execution progress value as the decay calculation base, combined with the updated risk particle flow intensity, to generate the antibody half-life. When the execution progress exceeds the threshold corresponding to the half-life, a clonal selection algorithm is forcibly triggered to recombine the decision antibody, generating a non-acute-angle shunt instruction to overwrite the original high-risk instruction. For example, if the initial execution progress is 40% and the threat level is downgraded, the system calculates the half-life based on the progress value and risk intensity. If the progress increases to 60% within the next three seconds, instruction recombination is triggered, the acute-angle baffle stops closing, and the system switches to a serpentine shunt channel.

[0058] Compared to traditional methods, this solution overcomes the limitations of rigid command execution. For the first time, the real-time coupling of progress values ​​and threat levels establishes a dynamic command optimization mechanism, ensuring that channel resource occupancy duration precisely matches the actual risk duration. The half-life model adaptively adjusts the reorganization timing based on risk intensity, avoiding residual risks caused by premature removal of controls in the early stages of downgrade. System resource utilization is improved to a level synchronized with the risk situation in real time, channel turnover efficiency approaches the theoretical optimal value, and the need for manual intervention is reduced to a fraction of that in traditional models.

[0059] The above technical solution establishes a dynamic instruction optimization mechanism. Using instruction execution progress as the decay base (progress value P < 50%), and combining it with the real-time risk particle flow intensity (R), the half-life (T ∝ P / R) is calculated. When the progress exceeds the threshold corresponding to T, the instruction is recombined. This ensures that before physical isolation is completed (the baffle is not fully closed), if the threat level degrades, non-acute angle diversion instructions can be switched in a timely manner, avoiding the ineffective occupation of channel resources. Testing has shown that this significantly reduces the excessive execution rate of high-risk instructions, and channel turnaround efficiency is also greatly improved.

[0060] In another technical solution, the urban rail transit security inspection system also includes a real-time video compensation unit, which continuously collects passenger flow video streams through a wide-angle camera array deployed on the side wall of the passage. The real-time video compensation unit performs the following operations: S5.1 Extract the passenger movement speed distribution matrix and direction change rate in the video stream using optical flow method; S5.2 When the standard deviation of the moving speed exceeds the preset speed tolerance and the rate of change of direction is greater than the direction threshold, an emergency topology pre-instruction is generated; S5.3 Input emergency topology pre-instructions in parallel to the immune decision center and the topology variable channel group; Among them, when the passenger flow dynamics prediction module outputs a delay, the immune decision center uses the emergency topology pre-instruction as an antigen input substitute to drive the clonal selection algorithm to generate decision antibodies containing channel topology instructions; after receiving the emergency topology pre-instruction, the topology variable channel group generates channel topology instructions according to the entropy value interval.

[0061] Existing systems typically employ static contingency plans when predictive data is delayed, such as forcibly switching to a low-precision scanning mode or requiring full manual takeover. During the New Year's Day passenger peak, a city transportation hub experienced a system downgrade to a full-channel wide-area screening mode due to a timeout in the prediction calculation. This resulted in three passengers carrying suspicious items not being identified in time, and they were only intercepted after manual review. Such contingency plans lack the ability to perceive instantaneous passenger flow dynamics and cannot distinguish between normal congestion and sudden risk clusters.

[0062] This solution constructs an emergency response link through a real-time video compensation unit. A wide-angle camera array deployed on the sidewalls of the passageway continuously collects passenger flow video streams, and an optical flow processing algorithm extracts the passenger movement speed distribution matrix and direction change rate. When the standard deviation of the movement speed exceeds a preset tolerance and the direction change rate exceeds a direction threshold, it is determined to be a sudden passenger flow disturbance, generating an emergency topology pre-instruction. This instruction includes passageway contraction ratio and diversion angle parameters, which are input in parallel to the emergency interface of the immune decision center and the execution interface of the topology variable passage group. When the prediction module output is delayed, the decision center uses the pre-instruction as an antigen input substitute to drive the clonal selection algorithm to generate decision antibodies. The passage group synchronously executes the instruction according to preset rules within the entropy value range; for example, high entropy values ​​trigger serpentine passageway contraction, while low entropy values ​​maintain straight-line diversion.

[0063] Compared to traditional contingency plans, this solution overcomes the limitations of passive degradation. Optical flow feature analysis enables emergency commands to accurately match actual passenger flow fluctuations, avoiding the risk of missed detections caused by system-wide degradation. Pre-commands drive the decision-making and execution layers in parallel, establishing a temporary response link within 200 milliseconds. The system automatically switches back to master control mode after the predicted data recovers, requiring no manual intervention throughout the process, achieving uninterrupted risk management continuity.

[0064] The above technical solution creates an emergency response channel. Based on the velocity standard deviation (>1.2 m / s²) and direction change rate (>45° / s) extracted using optical flow, it detects sudden changes in passenger flow. The generated emergency pre-instructions include channel width contraction rate and diversion angle parameters. When the predicted data is delayed, the immune decision center directly uses it as antigen input to generate antibody instructions. The channel group synchronously executes preset actions within the entropy value range (e.g., activating a serpentine channel when the entropy value >0.8). This enables the system to respond to sudden passenger flow surges within 200 ms, improving response speed compared to traditional systems.

[0065] In another technical solution, the real-time video compensation unit of the urban rail transit security inspection system further performs the following operations: S5.4 Continuously monitor the acquisition status of mobile communication base station signaling density distribution data and gesture skeleton key point vectors; S5.5 When the duration of missing signaling density distribution data exceeds the first tolerance or the invalid count of gesture skeletal keypoint vectors exceeds the second tolerance, activate the video compensation enhancement protocol and perform the following operations: The video stream collected by the wide-angle camera array is used to generate a passenger flow concentration distribution map through spatiotemporal slicing analysis, which replaces the missing signaling density distribution data; The passenger's upper limb movement trajectory is analyzed using a pose estimation algorithm to generate a simulated gesture skeletal key point vector; S5.7 Input the passenger flow concentration distribution map and the simulated gesture skeleton key point vector into the risk particle flow intensity generation step.

[0066] Existing systems typically respond to data source anomalies by triggering threshold alarms followed by manual intervention, or by simply interpolating to fill in missing values. In one smart station project, a mobile base station malfunction caused signaling data interruption. The system used historical averages to fill in the gaps, mistakenly identifying morning rush hour queues as abnormal clusters, and incorrectly triggering high-risk commands that resulted in a two-hour channel closure. More seriously, when the gesture recognition device is obstructed, the system defaults to a zero vector value, rendering the direction correction coefficient ineffective, and causing the risk path to deviate completely from the security inspector's actual handling direction.

[0067] This solution enhances the data resilience mechanism of the protocol through video compensation, continuously monitoring the acquisition status of signaling density data and gesture skeletal vectors. When signaling data loss persists beyond the first tolerance or the invalid gesture vector count exceeds the second tolerance, the compensation protocol is automatically activated. Spatiotemporal slicing analysis is performed on the wide-angle camera video stream, and passenger positions are overlaid along the time axis to generate a passenger flow clustering distribution map, replacing the invalid signaling density data. A posture estimation algorithm is used to analyze the motion trajectory of the passenger's upper limb joints, fitting and generating simulated gesture skeletal keypoint vectors to replace abnormal gesture data. The clustering distribution map is input to the abnormal region identification stage of the risk particle flow generation process, simulating the gesture vector input direction correction coefficient calculation process.

[0068] Compared to traditional methods, this solution overcomes the limitations of passive filling. The clustering distribution map generated by spatiotemporal slice analysis restores the spatiotemporal evolution characteristics of passenger flow, avoiding misjudging normal queuing as abnormal clustering. The skeletal vector fitted by posture estimation preserves the continuity of behavioral intent, ensuring that the dimension of human intervention is not lost due to equipment malfunction. The system maintains the integrity of risk prediction function even when the data source partially fails, with a false alarm rate close to the level of normal operation, and the decision reliability in failure scenarios is improved to a practical level.

[0069] The above technical solution achieves seamless switching in case of data failure. Spatiotemporal slicing analysis transforms the video stream into a spatiotemporal cube, which is then projected along the time axis to generate a passenger flow aggregation distribution map (resolution 0.5m / pixel), replacing the missing signaling density data. Pose estimation generates a 17-dimensional gesture skeleton vector through keypoint trajectory fitting, replacing the original data when the confidence level is >0.85. This ensures that the error rate of risk particle flow intensity calculation is effectively controlled in the event of signaling interruption or gesture acquisition failure, significantly enhancing the system's robustness.

[0070] In another technical solution, the urban rail transit security inspection system also includes a multimodal data redundancy engine, whose input end is connected to the real-time signaling data interface of the existing monitoring camera group and mobile communication base station in the station. The multimodal data redundancy engine performs the following operations: S6.1 Continuously monitor the equipment status codes of the ground pressure sensing network and 3D lidar. When the status code remains abnormal for a duration longer than the fault tolerance limit, activate the data reconstruction protocol. S6.2 After the data reconstruction protocol is activated, perform the following operations: a) Analyze the video streams from the surveillance camera group and generate a heat map of passenger flow density distribution using the background subtraction method; b) Extract the signaling hopping trajectory of the mobile communication base station and construct the passenger movement direction probability matrix; S6.3 Map the passenger flow density distribution heat map to simulated pressure distribution data, convert the passenger movement direction probability matrix into simulated motion vector field, and input it into the passenger flow dynamics prediction module to replace the original data source; S6.4 When the device status code returns to normal, it will automatically switch back to the original data acquisition mode.

[0071] Existing systems often employ threshold alarm shutdown strategies or switch to low-precision backup sensors to handle sensor failures. At the Shanghai Hongqiao Transportation Hub, a short circuit in the pressure sensor network forced the system to switch to a backup infrared counter. This counter only provided the number of people passing through, losing density distribution data, leading to the prediction module misjudging passageways as empty and failing to warn of the risk of stampedes caused by localized congestion. More seriously, when lidar malfunctions, the system fills motion data with fixed-direction vectors, causing the direction of the risk particle flow to deviate from the actual main passenger flow path.

[0072] This solution constructs a fault-tolerant system through a multimodal data redundancy engine: continuously monitoring the equipment status codes of the pressure sensing network and 3D LiDAR, activating a data reconstruction protocol when anomalies persist beyond the fault tolerance limit. It analyzes the video streams from existing station surveillance cameras, separating moving foreground pixels using background subtraction to generate a heatmap with grayscale values ​​proportional to passenger flow density; simultaneously extracting real-time signaling transition trajectories from mobile communication base stations, and constructing a passenger movement direction probability matrix based on a Markov chain model. The heatmap is linearly mapped to simulated pressure distribution data, and the direction probability matrix is ​​converted into a weighted motion vector field, which is then input into the passenger flow dynamics prediction module to replace the original fault data source. After equipment status recovery, it automatically switches back to primary sensing mode.

[0073] Compared to traditional shutdown strategies, this solution overcomes the limitations of data substitution distortion: the heat map generated by background difference retains the spatial gradient of passenger flow density, avoiding misjudging local high density as global idleness; the directional probability matrix constructed by signal jumps restores the actual movement trend, making the direction of risk particle flow match the real propagation path. The system maintains uninterrupted prediction function during core sensor failures, and the output accuracy meets basic risk management requirements, substantially overcoming hardware reliability shortcomings.

[0074] The above technical solution constructs a fault-tolerant system. Background subtraction is used to extract foreground pixels from the monitoring video to generate a heatmap (grayscale value proportional to density). The signaling jump trajectory is used to construct a direction probability matrix via a Markov chain. The heatmap is linearly mapped to pressure distribution data (grayscale value 0-255 corresponds to pressure 0-10 kPa), and the direction matrix is ​​converted into a motion vector field (probability > 0.7 is set as the principal direction). This maintains the continuity of data input to the prediction module during sensor failures, reducing system downtime.

[0075] In another technical solution, the urban rail transit security inspection system, when simultaneously acquiring mobile communication base station signaling density distribution data and gesture skeleton key point vector direction in step S1.1, activates the multi-source biometric fusion protocol when the effective coverage rate of the signaling density distribution data is less than the preset coverage threshold or the confidence level of the gesture skeleton key point vector is less than the confidence threshold. The multi-source biometric fusion protocol performs the following operations: The phase difference of passenger gait cycles is collected by an infrared thermal imaging array deployed at the top of the passage, generating a gait aggregation density distribution map, which is then normalized to replace the signaling density distribution data. The direction of passenger torso swing is analyzed by the micro-Doppler effect of millimeter-wave scanning array, and a biomechanical direction vector is generated. When the swing amplitude is greater than the sensitivity threshold, the gesture skeletal key point vector is replaced. The gait aggregation density distribution map is input into the abnormal aggregation region identification process in step S1.2, and the biomechanical direction vector is input into the second correction coefficient calculation process in step S1.3.

[0076] Existing systems lack effective redundancy mechanisms for handling anomalies in biometric data collection. At an international exhibition, a metal shield caused a disruption in motion signaling at a security checkpoint. The system abandoned risk calculations for that area and failed to identify abnormal clustering in the exhibit transport lane; contraband was later discovered manually. Similarly, during heavy rain, when security personnel's raincoats obscured their hand gesture devices, the system used a default positive vector, causing the risk path to deviate from the actual direction of the personnel's actions by more than 60 degrees, misleading the scanning array to focus on the wrong area.

[0077] This solution constructs an environmentally adaptive mechanism through a multi-source biometric fusion protocol: when the effective signaling coverage is insufficient or the gesture confidence is too low, the fusion protocol is activated. An infrared thermal imaging array at the top of the channel collects temperature fluctuations in the passenger's lower limb movements, and generates a normalized gait clustering density map through gait cycle phase difference analysis—regions with excessive phase difference are marked as high-density areas, replacing the missing signaling density data input to the abnormal clustering identification module. A millimeter-wave scanning array simultaneously captures the micro-Doppler frequency shift reflected from the passenger's torso; when the swing amplitude meets the standard, the swing direction vector is analyzed, replacing the occluded gesture skeletal key points input for direction correction calculation.

[0078] Compared to traditional abandonment strategies, this solution overcomes environmental constraints. Gait thermal imaging reconstructs passenger flow patterns through physiological rhythm characteristics, eliminating signal blind spots and monitoring vacuums. Millimeter-wave micro-Doppler penetratingly captures the intention of torso swaying, maintaining continuity in the dimension of human intervention. The system maintains complete risk prediction functionality even in harsh acquisition environments, with output reliability reaching levels comparable to conventional operating conditions, effectively bridging the security shortcomings in special scenarios.

[0079] The above technical solution overcomes the limitations of biometric data acquisition. Gait phase difference is calculated using the temperature fluctuation period of the lower limbs from infrared thermography to generate a normalized cluster density map (phase difference > 15° is designated as a high-density area). Millimeter-wave micro-Doppler frequency shift analysis of trunk sway (frequency shift > 50Hz when amplitude > 15°) generates a biomechanical vector to replace gesture data. In scenarios with insufficient signaling coverage or gesture occlusion, it improves the accuracy of abnormal area recognition and reduces the error rate of direction correction coefficient calculation.

[0080] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0081] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details.

Claims

1. A security inspection system for urban rail transit, characterized in that, include: The passenger flow dynamics prediction module constructs a passenger flow micro-environment model based on the collected passenger flow pressure distribution and motion vector field, and outputs congestion entropy value and risk particle flow in real time; the congestion entropy value is a scalar indicator that quantifies the degree of regional congestion, and the risk particle flow is a vector field that characterizes the direction and intensity of risk propagation. An adaptive millimeter-wave scanning array receives congestion entropy values ​​and risk particle streams. When the intensity of the risk particle stream exceeds a predetermined intensity threshold, a moving focusing scanning mode is activated. When the risk particle flow intensity is less than or equal to the predetermined intensity threshold and the congestion entropy value is greater than the predetermined entropy threshold, the wide-area screening mode is activated and a basic threat index is generated. When a passenger's basic threat index is found to be greater than the safety threshold, the passenger is marked as a target to be re-inspected. If both the risk particle flow intensity and the congestion entropy value are less than or equal to the threshold, the wide-area screening mode is executed. The feature analysis module, in the mobile focusing scanning mode, receives and parses millimeter-wave scanning data, extracts the outline features of the items carried by the passenger being focused, compares the outline features with a pre-stored contraband feature database, and generates the passenger's real-time threat level. When a target identifier to be re-inspected is received in wide-area screening mode, a lightweight analysis model is activated to analyze the millimeter-wave data of the target to be re-inspected and generate a threat assessment report of the target to be re-inspected. The immune decision center receives congestion entropy values, real-time threat levels, and threat assessment reports of targets to be re-examined. It dynamically generates decision antibodies through a cloning selection algorithm. These decision antibodies contain channel topology instructions and device parameter sets. The generation rules for channel topology instructions include: if the threat information is a high-risk real-time threat level, then instructions containing a 30° acute angle fluid splitting are forcibly generated. If the threat information is a non-high-risk real-time threat level, then a corresponding channel topology instruction is generated based on the predetermined entropy value range where the congestion entropy value is located; when the target identifier to be re-inspected and the corresponding target threat assessment report to be re-inspected are received, a channel topology instruction pointing to the secondary diversion area is generated.

2. The urban rail transit security inspection system as described in claim 1, characterized in that, Also includes: The variable topology channel group receives and executes channel topology instructions and equipment parameter sets, dynamically adjusting the channel physical structure; based on instructions including 30° acute-angle flow dividers, it forms acute-angle flow divider structures at corresponding positions; based on instructions pointing to secondary flow dividers, it guides the target to be re-inspected to its preset secondary flow divider zone; within the secondary flow divider zone, it further constructs the channel structure based on channel topology instructions generated from the threat assessment report for the target to be re-inspected; when the threat assessment report indicates high risk, it constructs a π-shaped aggregate channel to guide the target to be re-inspected to the re-inspection zone; it collects real-time data on changes in pedestrian flow pressure within the channel and feeds it back to the passenger flow dynamics prediction module for calibrating the passenger flow microenvironment dynamics model.

3. The urban rail transit security inspection system as described in claim 2, characterized in that, The intensity value of the risk particle stream is dynamically generated through the following steps: S1.1 Real-time access to the risk path spatial distribution data output by the passenger flow dynamics prediction module, synchronously acquiring the mobile device signaling density distribution data provided by the mobile communication base station, and the gesture skeleton key point vector direction captured by the security inspector's wearable device; S1.

2. Based on mobile device signaling density distribution data, identify the boundary coordinates of abnormal clustering areas; based on risk path spatial distribution data, extract the path vector of the main diffusion direction; S1.3 Calculate the percentage of overlap between the boundary coordinates of the abnormal clustering area and the spatial distribution area of ​​the risk path, and use it as the first correction coefficient; Calculate the cosine of the angle between the direction of the gesture skeleton keypoint vector and the direction of the path vector of the main diffusion direction, and use it as the second correction coefficient; S1.

4. The frequency of occurrence of historical prohibited items stored in the security check event database is used as the basic weight. Combined with the first correction coefficient and the second correction coefficient, it is input into the pre-trained risk propagation tree model. The risk propagation tree model is generated by pre-training based on the spatiotemporal distribution pattern of the historical security check event database. S1.5 The risk propagation tree model outputs the updated risk particle flow intensity value and feeds it back to the passenger flow dynamics prediction module.

4. The urban rail transit security inspection system as described in claim 2, characterized in that, Also includes: The dynamic path calibration engine is connected to the initial path output of the risk particle flow of the passenger flow dynamics prediction module at its input end, and is connected in parallel to the multispectral interference sensing array deployed on the tunnel arch. A multispectral interference sensing array simultaneously acquires the physiological agitation index in the thermal infrared band and the motion change trajectory in the visible light band; the physiological agitation index refers to the quantitative indicator output by the rate of temperature rise of the human skin, which is acquired by a thermal infrared camera deployed on the arch of the passage. When the physiological agitation index exceeds the preset agitation threshold and lasts for more than 5 seconds, or when the cumulative deviation of the azimuth angle between the sudden movement trajectory and the initial path is greater than or equal to 30°, the nonlinear compensator is activated. Once the nonlinear compensator is activated, it performs the following operations: S3.1 Map the physiological agitation index to a path offset probability density function, and generate the first compensation vector through a convolution kernel; S3.2 Decompose the tangential and normal acceleration components of the trajectory of sudden motion change, and generate a second compensation vector with the amplitude of the normal component as the weight; S3.3, merge the first compensation vector and the second compensation vector into the initial path of the risk particle flow, and output the anti-disturbance calibration path to the risk path input of the adaptive millimeter-wave scanning array.

5. The urban rail transit security inspection system as described in claim 4, characterized in that, After the immune decision-making center generates decision antibodies, continuous antibody activity monitoring is performed: S4.

1. Real-time acquisition of threat level update flags and instruction execution progress values ​​of the topology variable channel group from the feature analysis module; S4.2 When the progress value of the channel group executing the high-risk acute angle fluid splitting command is less than 50%, if the threat level update flag is switched to the medium-low risk state, the antibody apoptosis protocol is activated. S4.3 The antibody apoptosis protocol uses the instruction execution progress value as the decay base and combines it with the updated risk particle flux intensity value to generate the antibody half-life; S4.4 When the instruction execution progress value exceeds the progress threshold corresponding to the antibody half-life, the clone selection algorithm is forcibly triggered to recombinate the decision antibody. The recombination rule uses non-acute angle diversion instructions generated by the entropy value interval to cover the original high-risk instructions.

6. The urban rail transit security inspection system as described in claim 3, characterized in that, It also includes a real-time video compensation unit, which continuously collects passenger flow video streams through a wide-angle camera array deployed on the side wall of the passageway; The real-time video compensation unit performs the following operations: S5.1 Extract the passenger movement speed distribution matrix and direction change rate in the video stream using optical flow method; S5.2 When the standard deviation of the moving speed exceeds the preset speed tolerance and the rate of change of direction is greater than the direction threshold, an emergency topology pre-instruction is generated; S5.3 Input emergency topology pre-instructions in parallel to the immune decision center and the topology variable channel group; Among them, when the passenger flow dynamics prediction module outputs a delay, the immune decision center uses the emergency topology pre-instruction as an antigen input substitute to drive the clonal selection algorithm to generate decision antibodies containing channel topology instructions; after receiving the emergency topology pre-instruction, the topology variable channel group generates channel topology instructions according to the entropy value interval.

7. The urban rail transit security inspection system as described in claim 6, characterized in that, The real-time video compensation unit further performs the following operations: S5.4 Continuously monitor the acquisition status of mobile communication base station signaling density distribution data and gesture skeleton key point vectors; S5.5 When the duration of missing signaling density distribution data exceeds the first tolerance or the invalid count of gesture skeletal keypoint vectors exceeds the second tolerance, activate the video compensation enhancement protocol and perform the following operations: The video stream collected by the wide-angle camera array is used to generate a passenger flow concentration distribution map through spatiotemporal slicing analysis, which replaces the missing signaling density distribution data; The passenger's upper limb movement trajectory is analyzed using a pose estimation algorithm to generate a simulated gesture skeletal key point vector; S5.6 Input the passenger flow concentration distribution map and the simulated gesture skeleton key point vector into the risk particle flow intensity generation step.

8. The urban rail transit security inspection system as described in claim 1, characterized in that, It also includes a multimodal data redundancy engine, whose input end is connected to the real-time signaling data interface of the station's existing surveillance camera group and mobile communication base station; The multimodal data redundancy engine performs the following operations: S6.1 Continuously monitor the equipment status codes of the ground pressure sensing network and 3D lidar. When the status code remains abnormal for a duration longer than the fault tolerance limit, activate the data reconstruction protocol. S6.2 After the data reconstruction protocol is activated, perform the following operations: a) Analyze the video streams from the surveillance camera group and generate a heat map of passenger flow density distribution using the background subtraction method; b) Extract the signaling hopping trajectory of the mobile communication base station and construct the passenger movement direction probability matrix; S6.3 Map the passenger flow density distribution heat map to simulated pressure distribution data, convert the passenger movement direction probability matrix into simulated motion vector field, and input it into the passenger flow dynamics prediction module to replace the original data source; S6.4 When the device status code returns to normal, it will automatically switch back to the original data acquisition mode.

9. The urban rail transit security inspection system as described in claim 3, characterized in that, In step S1.1, when the signaling density distribution data of the mobile communication base station and the direction of the gesture skeleton key point vector are acquired synchronously, if the effective coverage rate of the signaling density distribution data is less than the preset coverage threshold or the confidence level of the gesture skeleton key point vector is less than the confidence threshold, the multi-source biometric fusion protocol is activated. The multi-source biometric fusion protocol performs the following operations: The phase difference of passenger gait cycles is collected by an infrared thermal imaging array deployed at the top of the passage, generating a gait aggregation density distribution map, which is then normalized to replace the signaling density distribution data. The direction of passenger torso swing is analyzed by the micro-Doppler effect of millimeter-wave scanning array, and a biomechanical direction vector is generated. When the swing amplitude is greater than the sensitivity threshold, the gesture skeletal key point vector is replaced. The gait aggregation density distribution map is input into the abnormal aggregation region identification process in step S1.2, and the biomechanical direction vector is input into the second correction coefficient calculation process in step S1.3.

Images