A method and system for classifying travel users

Abstract

The present application relates to the technical field of data processing, in particular to a travel user classification method and system. The method comprises: obtaining pre-security multi-source real-time detection data, historical data and space-time trajectory data of a target user; generating a pre-security multi-dimensional data set based on the pre-security multi-source real-time detection data; inputting the pre-security multi-dimensional data set and the historical data into a multi-level weighted calculation model to calculate a fusion characteristic value; determining an initial processing channel level according to the fusion characteristic value; when the initial processing channel level is a first processing channel level, adjusting the initial processing channel level based on the space-time trajectory data, and taking the adjusted processing channel level as a classification result. The present application constructs a two-stage architecture of "first access, then calibration", dynamically calibrates the real-time trajectory of the first processing channel user, and realizes the adaptive matching of security inspection intensity while ensuring the traffic efficiency.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for classifying travel users. Background Technology

[0002] Currently, differentiated processing models based on intelligent data analysis have emerged in the field of civil aviation security inspection. For example, Chinese patent application CN120105064B discloses an intelligent analysis method and system for airport security inspection data. It acquires multi-source security inspection data (including baggage scanning images, behavioral monitoring videos, identity verification data, etc.), performs multimodal data cleaning, dangerous feature extraction, and spatial feature extraction to generate security inspection characteristic sequence data. Then, through behavioral mutation detection and cross-modal feature correlation, it assesses the threat level of mutated behavior and cross-modal threat level to assist security inspection decisions.

[0003] However, the above-mentioned existing technical solutions still have the following technical problems:

[0004] First, this solution primarily focuses on threat identification of baggage and anomaly detection, outputting "threat level" and "danger data" to assist security personnel in decision-making, rather than classifying individual passengers into different access channels. The solution fails to establish a multi-level weighted scoring system based on passenger risk levels, thus failing to achieve the technical effect of dynamically diverting passengers to different security checkpoints (such as fast lanes, standard lanes, and enhanced lanes).

[0005] Second, the scheme's processing of passenger spatiotemporal trajectory data during behavioral analysis is relatively simplistic (e.g., using mutation detection to mark abnormal stops or reverse travel), failing to introduce a segmented dynamic clustering comparison model. This means it doesn't consider the dynamic impact of the time interval between the passenger's departure and flight departure on the normal trajectory range, nor does it establish an adaptive calibration mechanism for different cluster radius thresholds corresponding to different time periods. Therefore, this scheme is prone to generating high false alarm or false negative rates when judging trajectory anomalies.

[0006] Third, the scheme does not introduce facial expression recognition as a dynamic correction factor for trajectory deviation thresholds, nor does it set up a hierarchical rapid exit mechanism and a parallel computing optimization architecture, making it difficult to meet the airport's real-time requirements while achieving refined passenger classification.

[0007] Therefore, how to integrate multi-source real-time detection data with historical data, construct a dynamic classification mechanism for individual passenger differences, and introduce segmented trajectory clustering calibration and facial expression correction coefficients to achieve secondary dynamic verification of users undergoing rapid security checks are technical problems that urgently need to be solved in this field. Summary of the Invention

[0008] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows:

[0009] According to a first aspect of the present invention, a method for classifying travel users is provided, comprising:

[0010] Acquire multi-source real-time detection data, historical data, and spatiotemporal trajectory data of the target user for pre-security checks.

[0011] Based on the pre-security check multi-source real-time detection data, a pre-security check multidimensional dataset for the target user is generated.

[0012] The pre-security check multidimensional dataset and the historical data are input into a multi-level weighted calculation model to calculate the fusion feature value of the target user.

[0013] Based on the fusion feature value, the initial processing channel level corresponding to the target user is determined, and the initial processing channel level includes a first processing channel level corresponding to the fusion feature value being greater than or equal to a first threshold.

[0014] When the initial processing channel level is the first processing channel level, the initial processing channel level is adjusted based on the spatiotemporal trajectory data of the target user, and the adjusted processing channel level is used as the classification result of the target user.

[0015] According to a second aspect of the present invention, a travel user classification system is provided, comprising:

[0016] One or more processors;

[0017] Memory, used to store one or more programs;

[0018] When the one or more programs are executed by the one or more processors, the one or more processors implement the travel user classification method as described in the first aspect.

[0019] This invention constructs a two-stage hierarchical classification architecture of "first admission, then calibration": In the first stage, multi-level weighted fusion is performed based on pre-security inspection multi-source real-time detection data and historical data to calculate the fusion feature value of the target user. Users whose fusion feature value is greater than or equal to a first threshold are initially screened to the first processing channel level, realizing rapid diversion of low-abnormal users based on static and quasi-static features; In the second stage, for users already assigned to the first processing channel level, their continuous spatiotemporal trajectory data is called for dynamic trajectory calibration. Through segmented clustering comparison and quantitative mapping of the deviation degree in feature space, the trajectory segment aggregation abnormality level is converted into a channel upgrade instruction, forming a closed-loop control loop of "admission-monitoring-feedback-adjustment".

[0020] This invention overcomes the technical shortcomings of traditional differentiated security checks, which rely on static binding between single-score and channel level, leading to an inability to adaptively respond to real-time behavioral abrupt changes. It achieves a nonlinear match between security check intensity and high-dimensional spatiotemporal behavioral characteristics, constructing a posterior anomaly compensation mechanism based on real-time trajectories while ensuring passage efficiency. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of the invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating a method for classifying travel users according to an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0025] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. A process can be terminated when its operation is complete, but it may also have additional steps not included in the figures. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0026] First Embodiment

[0027] This invention provides a method for classifying travelers, applied to the security check process at civil aviation airports. For example... Figure 1 As shown, the method includes:

[0028] S100 acquires multi-source real-time detection data, historical data, and spatiotemporal trajectory data of the target user for pre-security checks.

[0029] The target user refers to an individual passenger who is to be classified into different security checkpoint levels.

[0030] Specifically, multi-source real-time detection data is collected through a pre-security check system deployed at the terminal entrance. The pre-security check system includes: a rapid, non-contact detection pre-security check gate, a terahertz body scanner, a baggage X-ray machine, and a system management platform.

[0031] The rapid, non-intrusive pre-security check gate integrates a trace explosive detector, a neutron detector, a gamma ray detector, and a combustible volatile substance detector. It is used to simultaneously conduct comprehensive security detection during passenger identity verification, obtaining identity verification results and multi-source security detection results. The terahertz body scanner is used to scan prohibited items carried by personnel, obtaining body security check results. The baggage X-ray machine is used to scan prohibited items in baggage, obtaining baggage security check results. The system management platform is used to record passage judgment and control information, multi-channel equipment management information, and pre-security check original alarm data.

[0032] The aforementioned pre-security inspection multi-source real-time detection data includes, but is not limited to, the following six types of raw data:

[0033] Identity verification data: person-document matching degree, document validity, basic identity identifiers;

[0034] Multi-source security detection data: Explosive detection results, radioactive material detection results (including neutron / gamma ray detection results), and combustible volatile material detection results.

[0035] Data related to personal security checks: Detection of prohibited items on terahertz human bodies, and suspicious locations;

[0036] Baggage security inspection data: baggage X-ray imaging results, whether baggage is carried, and results of suspicious items in baggage;

[0037] Behavioral characteristic data: walking gait, dwell time, trajectory, abnormal loitering or inspection avoidance behavior;

[0038] System alarms and discrimination data: security anomaly alarms, behavior anomaly alarms, and access compliance judgment records.

[0039] The above six types of data are raw data directly output by the pre-security check system, and have not yet undergone structured fusion and feature encoding. Subsequent steps will involve cleaning, feature extraction, and fusion encoding of this raw data to generate a pre-security check multidimensional dataset.

[0040] The spatiotemporal trajectory data is a continuous sequence of spatiotemporal trajectory points of the target user within a target time interval. Each trajectory point includes a spatial x-coordinate, a spatial y-coordinate, and a timestamp. The sequence of trajectory points is arranged in chronological order and is used to characterize the continuous movement path of the target user within the terminal. The spatial x-coordinate and spatial y-coordinate are determined based on a pre-established two-dimensional plane coordinate system of the airport terminal. This coordinate system takes a preset reference point within the terminal (such as the center point of a security checkpoint entrance or the terminal's geographical reference point) as its origin and preset directions (such as the X-axis parallel to the main passage of the terminal and the Y-axis perpendicular to the main passage) as its coordinate axis directions.

[0041] The target time interval begins at the first moment after the target user enters the airport or completes pre-security check, and ends at the second moment when the target user enters the waiting area or the time remaining until the expected event (such as flight departure time) reaches a preset threshold. The preset threshold is pre-configured based on airport operating parameters or flight type, and defines the time boundary at which the target user is no longer subject to dynamic trajectory calibration constraints; for example, the preset threshold can be set to 15 minutes, 20 minutes, or 30 minutes. The specific start and end times can be adjusted according to the actual application scenario.

[0042] The spatiotemporal trajectory data can be collected by positioning base stations (such as UWB ultra-wideband positioning systems) or video surveillance systems deployed in the terminal building. During collection, the coordinate values ​​of each trajectory point are calibrated or transformed based on the two-dimensional plane coordinate system.

[0043] It should be noted that the trajectory lines, dwell time, and abnormal loitering or evasion behaviors in the behavioral feature data are derived from real-time video analysis and multimodal intelligent perception of the pre-security inspection system, used for immediate behavior identification; while the spatiotemporal trajectory data comes from an independent positioning system (such as UWB or video surveillance), providing higher precision and higher frequency continuous trajectory point sequences for subsequent segmented clustering comparison. The two types of data have different sources, different precision, and different uses, and complement each other.

[0044] The historical data includes: historical security record data, historical flight record data, and flight attribute data. Historical security record data can be accessed through a data interface to an authorized historical security data platform, using the target user's unique identifier (such as ID card number or passport number) as the query key to retrieve historical security records, preset-identified personnel identifiers, and records of dishonesty. Encrypted transmission protocols are used during data acquisition to ensure data security. Historical flight record data can be accessed through a data interface to the civil aviation departure system or passenger service system, using the target user's unique identifier as the query key to retrieve data on flight frequency, security violation records, route stability, and ticketing behavior over the past three years. Flight attribute data can be accessed through a data interface to the flight information management system to retrieve the current flight's flight time, route type (domestic / international, whether it is a specific route area), holiday markings, and whether it is a temporary / charter flight.

[0045] The aforementioned historical data can be obtained before the target user arrives at the airport through offline batch synchronization or real-time on-demand querying, and cached on the local server to reduce the overhead of repeated queries.

[0046] Through the above steps, complete input data is obtained for subsequent multi-level weighted calculations and dynamic trajectory calibration.

[0047] S200, Based on the pre-security check multi-source real-time detection data, generate the pre-security check multidimensional dataset of the target user.

[0048] Specifically, the six types of pre-security inspection multi-source real-time detection data obtained in step S100 are cleaned, feature extracted, standardized and encoded, and multi-dimensionally fused to generate a pre-security inspection multidimensional dataset that can be directly input into a multi-level weighted calculation model.

[0049] (a) Data cleaning

[0050] Outlier detection and missing value handling are performed on various types of raw detection data:

[0051] For the original identity verification data: remove outliers where the identity document matching degree exceeds the range of [0,1], and fill and mark the missing document validity field with the default value "invalid".

[0052] For raw data from multi-source security detection: invalid records generated by sensor self-test errors are deleted, and median filtering is applied to the instantaneous noise values ​​in the detection results of explosives, radioactive materials, etc.

[0053] For raw data of personal security checks: false alarms in the detection of prohibited items on the terahertz body (such as multiple consecutive frames of detection but no corresponding item) are eliminated by time window voting method.

[0054] For raw baggage security inspection data: Based on the baggage X-ray imaging results, filter out artifact labels caused by baggage obstruction or overlap.

[0055] For the raw behavioral characteristic data: interpolation is performed to complete segments with insufficient gait sampling rates, and outliers in dwell time (such as those exceeding the maximum possible dwell time in the terminal) are truncated.

[0056] For the original system alarm data: deduplicatize and merge consecutive alarms of the same type (e.g., merge repeated security anomaly alarms within 10 seconds into one valid alarm).

[0057] It should be noted that the data cleaning operations involved in this step (including but not limited to median filtering, time window voting method, interpolation completion, outlier truncation, data deduplication and merging) are all implemented using existing technologies known in the field. Those skilled in the art can select appropriate parameters and implement them according to the data characteristics, so the specific algorithm details will not be elaborated further.

[0058] (II) Feature Extraction and Quantization

[0059] Extract quantifiable feature indicators from the cleaned data:

[0060] Identity verification features: the degree of matching between person and document is directly used as a continuous value between 0 and 1; document validity is converted into a binary variable (valid = 1, invalid = 0); basic identity identifiers (age, gender) are mapped to age group codes (for example, 0-12 years old corresponds to code 1, 13-17 years old corresponds to code 2, 18-60 years old corresponds to code 3, and 60 years old and above corresponds to code 4), and gender codes use 1 for male and 0 for female.

[0061] Multi-source security detection features: Explosives, radioactive materials, combustible volatiles, and neutron / gamma detection results are quantified as binary variables (negative=0, positive=1), and a positive result is set to directly trigger a high-risk flag.

[0062] Features of personal security checks: The detection status of terahertz human contraband is converted into the number of detected items (0, 1, ≥2 are mapped to 0, 1, 2 respectively) and the highest risk level (general contraband = 1, controlled items = 2); the number of suspicious locations is normalized to the [0,1] interval.

[0063] Baggage security inspection features: Baggage X-ray imaging results are coded as normal=0, suspected abnormal=1, confirmed abnormal=2; whether baggage is carried is binarized (yes=1, no=0); baggage suspicious items are identified by category and the maximum value is taken.

[0064] Behavioral characteristics: Gait stability score (0~1) is calculated by acceleration variance; Dwell time is normalized by dividing by a baseline time (e.g., 300 seconds); Trajectory is calculated by the ratio of path length to straight-line distance to determine tortuosity; Abnormal loitering / avoidance of inspection behavior is quantified by the labels output by the pre-security inspection system's built-in rules (none=0, suspicious=1, abnormal=2).

[0065] System alarm characteristics: security anomaly alarms, behavior anomaly alarms, and access compliance judgments are converted into binary or ternary variables, and the alarm count and the time difference between the most recent alarm time and the current time are recorded.

[0066] (III) Standardized Coding and Multidimensional Fusion

[0067] All extracted feature metrics are standardized to eliminate the influence of different units of measurement. Standardization can be achieved using Z-score normalization or Min-Max normalization, with the corresponding normalization parameters pre-calculated and stored based on historical data distribution. The standardized feature metrics are then concatenated into a fixed-dimensional feature vector according to a pre-defined feature order, serving as the basic representation of the pre-security inspection multidimensional dataset. Simultaneously, the original labels and corresponding timestamps of each feature are retained, forming a structured dataset with time-series labels.

[0068] (iv) Output pre-security check multidimensional dataset

[0069] The final pre-security check cube includes:

[0070] A high-dimensional feature vector (the dimension is equal to the total number of selected feature indicators, such as the seven secondary indicators under the pre-security check category among the five primary indicators, and some indicators in the passenger identity and behavior feature category).

[0071] An auxiliary metadata dictionary records information such as data acquisition time, device number, and data integrity flags.

[0072] Optionally, positive flags from multi-source security detection can be stored independently for direct reading by the fast exit mechanism in the multi-level weighted calculation model.

[0073] The aforementioned pre-security check multidimensional dataset, together with the historical data obtained in step S100, is input into the subsequent multi-level weighted scoring model.

[0074] S300, input the pre-security check multidimensional dataset and the historical data into the multi-level weighted calculation model to calculate the fusion feature value of the target user.

[0075] The aforementioned anomaly category refers to the anomaly category that the target user needs to be directly assigned to the fourth processing channel (i.e., the manual review channel) in subsequent processing. Specifically, it includes situations such as explosion-proof detection anomaly, combustible volatile material detection anomaly, radioactive material detection anomaly, carrying controlled prohibited items, existence of anomaly records, or belonging to a person with a preset identification, which trigger the fast exit mechanism.

[0076] The multi-level weighted calculation model contains five feature layers, and each feature layer and its secondary features are assigned preset weight values ​​and preset quantization rules.

[0077] (I) First Feature Layer: Pre-security Inspection Features

[0078] This feature layer has a first preset weight value (e.g., 40%), which is used to quantitatively assess the security status of the target user in the pre-security check phase.

[0079] Feature Design Principles Explanation: The seven secondary features included in this feature layer (identity verification, explosive detection, flammable volatile substance detection, radioactive material detection, prohibited items detection, baggage X-ray scanning, and baggage carrying status) are all core security detection data collected in real time by the pre-security inspection system. Among them, explosives, radioactive materials, and flammable volatile substances are considered major security threats, and the routine scoring process must be terminated immediately upon detection; identity verification is the basic verification of identity authenticity; prohibited items detection and baggage scanning directly reflect the security status of passengers' carried items; and the number of bags serves as an auxiliary indicator reflecting the complexity.

[0080] The quantification rules and scoring logic for each secondary feature are as follows:

[0081] Identity verification feature: A second preset weight value (e.g., 6%) is set. When the verification is successful and the identity matches the document, a first preset positive value (e.g., +10 points) is assigned; when the verification fails or the identity does not match the document, a first preset negative value (e.g., -10 points) is assigned.

[0082] Explosion-proof detection feature: A third preset weight value (e.g., 8%) is set. When there are no traces of explosives, a second preset positive value (e.g., +15 points) is assigned; when there are traces of explosives, a second preset negative value (e.g., -15 points) is assigned, and the anomaly category determination is triggered directly, stopping all subsequent calculations.

[0083] Combustible volatile matter detection feature: A fourth preset weight value (e.g., 5%) is set. When there are no combustible volatile matter, a third preset positive value (e.g., +10 points) is assigned; when there are combustible volatile matter, a third preset negative value (e.g., -10 points) is assigned, and an anomaly category determination is triggered directly.

[0084] Radioactive material detection characteristics: A fifth preset weight value (e.g., 5%) is set. When there is no radioactive material, a fourth preset positive value (e.g., +10 points) is assigned; when there is radioactive material, a fourth preset negative value (e.g., -10 points) is assigned, and an anomaly category determination is triggered directly.

[0085] Contraband detection features: A sixth preset weight value (e.g., 5%) is set. A fifth preset positive value (e.g., +10 points) is assigned when no contraband is found; a fifth preset negative value (e.g., -10 points) is assigned when contraband is found. If the contraband is classified as controlled, an anomaly category determination is triggered directly.

[0086] Baggage X-ray scan features: A seventh preset weight value (e.g., 6%) is set. When there are no abnormalities in the baggage scan, a sixth preset positive value (e.g., +12 points) is assigned; when there are suspected prohibited items, a first intermediate negative value (e.g., -6 points) is assigned; when there are confirmed prohibited items, a sixth preset negative value (e.g., -12 points) is assigned.

[0087] Luggage carrying characteristics: An eighth preset weight value (e.g., 5%) is set. When there is no luggage or only a small amount of carry-on luggage, a seventh preset positive value is assigned (e.g., +8 points); when the amount of luggage is unusually large or the appearance of the luggage is suspicious, a seventh preset negative value is assigned (e.g., -3 points).

[0088] Scoring logic explanation: Identity verification is the foundation of identity authenticity. Failure to verify directly indicates that the identity is suspicious, so a negative score is set; successful verification increases the trust level.

[0089] Explosives are the highest level of security threat. If a positive result is detected, the routine scoring process must be terminated immediately and the case transferred directly to the manual review channel.

[0090] Combustible volatile substances (such as gasoline, alcohol, etc.) are flammable and explosive, posing a major safety hazard. A positive test result will also be subject to a veto.

[0091] Radioactive materials may be involved in nuclear safety or public hazard events, and must be isolated and handled immediately upon detection;

[0092] Common prohibited items (such as lighters and excessive liquids) can be handled through the regular process, resulting in point deductions but not termination of the scoring; controlled prohibited items directly trigger high-risk status.

[0093] The reliability of baggage scan results is divided into three levels: no abnormalities is the best; suspected abnormalities will result in a deduction of some points but leave room for further inspection; confirmed abnormalities will result in a significant deduction of points.

[0094] The quantity and complexity of luggage affect the efficiency of security checks. Excessive luggage or luggage with unusual shapes (such as being bundled or covered) may conceal prohibited items. Slight deductions are set to reflect potential risks.

[0095] (ii) Second feature layer: User identity and behavioral characteristics

[0096] This feature layer has a ninth preset weight value (e.g., 25%), which is used to evaluate the target user's basic identity attributes and real-time behavioral performance.

[0097] Feature Design Principles Explained: This feature layer includes five features: age, gender, facial expression, movement trajectory, and passage speed. It assesses potential risks from two dimensions: identity attributes and real-time behavioral performance. Statistical data shows that risk probabilities differ across age groups; facial expressions are identified through AI to recognize emotional states, and abnormal emotions may be related to carrying prohibited items or an intention to evade inspection; movement trajectory and passage speed reflect real-time behavioral patterns of passengers in the security checkpoint area, and abnormal paths or speed changes may indicate an intention to evade inspection.

[0098] The quantification rules and scoring logic for each secondary feature are as follows:

[0099] Age structure characteristics: A tenth preset weight value (e.g., 4%) is set. Different preset scores are assigned to different age groups. For example, children (0-12 years old) are assigned the first age score (e.g., +8 points), teenagers (13-17 years old) are assigned the second age score (e.g., +5 points), adults (18-60 years old) are assigned the third age score (e.g., +4 points), and the elderly (60 years and older) are assigned the fourth age score (e.g., +6 points).

[0100] Gender characteristic: An eleventh preset weight value is set (e.g., 2%). Both males and females are assigned the same preset score (e.g., +3 points).

[0101] Facial expression analysis features: A twelfth preset weight value (e.g., 3%) is set. Based on facial images collected by the pre-security inspection system, facial motion unit features are extracted through a deep learning model and compared with a preset expression template library. When the activation level of facial motion units is below the first preset threshold (e.g., 0.3) and no tension or fear-related feature units are activated, the expression is judged as natural and assigned a first expression score (e.g., +6 points). When the activation level of motion units around the eyes (e.g., AU1, AU2, AU4) is between the first and second preset thresholds (e.g., 0.3-0.6), or when the corners of the mouth are slightly pursed, the expression is judged as slightly tense and assigned a second expression score (e.g., +3 points). When the activation level of motion units exceeds the second preset threshold (e.g., 0.6), accompanied by auxiliary features such as sweating on the forehead and frequent blinking, the expression is judged as obviously tense and assigned a third expression score (e.g., -2 points). When the eyes are wide open (AU5) and the eyebrows are raised (AU1+AU2) in combination, or when a fear expression feature is detected and the confidence level exceeds the third preset threshold (e.g., 0.7), the expression is judged as fear or abnormal and assigned a fourth expression score (e.g., -6 points).

[0102] Movement trajectory characteristics: A thirteenth preset weight value (e.g., 3%) is set. Based on the real-time positioning or video analysis of passengers during the pre-security check process, the actual movement path of passengers is compared with the preset standard passage path. When the cumulative deviation distance between the actual path and the standard path is less than the first deviation threshold (e.g., 2 meters) and there is no turning back or detour behavior, the trajectory is judged as normal and a first trajectory score (e.g., +5 points) is assigned; when the cumulative deviation distance is between the first and second deviation thresholds (e.g., 2-5 meters), or when there is one unnecessary stop (the stop time is shorter than the first stop threshold, such as 10 seconds), it is judged as a slight abnormality and a second trajectory score (e.g., +2 points) is assigned; when the cumulative deviation distance exceeds the second deviation threshold (e.g., 5 meters), or when there is loitering behavior (repeatedly moving in unnecessary areas), or when there is inspection avoidance behavior (actively bypassing the security equipment area), it is judged as a significant abnormality and a third trajectory score (e.g., -5 points) is assigned.

[0103] Traffic speed characteristics: A fourteenth preset weight value (e.g., 2%) is set. It is determined based on the average moving speed of passengers from passing through the pre-security check gate to entering the waiting area. When the moving speed is between the first speed threshold and the second speed threshold (e.g., 0.8-1.5 m / s), it is judged as normal traffic speed and assigned a first speed score (e.g., +3 points); when the moving speed is lower than the first speed threshold (e.g., 0.8 m / s) and not due to reasonable reasons such as queuing, the elderly, or children, it is judged as too slow; when the moving speed is higher than the second speed threshold (e.g., 1.5 m / s) and is in a running or brisk walking state, it is judged as too fast. Both too slow and too fast are assigned a second speed score (e.g., +1 point); when the stationary stay time in the non-queuing area exceeds the preset stay threshold (e.g., 5 minutes), it is judged as stay time exceeding the limit and assigned a third speed score (e.g., -3 points).

[0104] Scoring logic explanation:

[0105] Statistics show that the proportion of adult passengers carrying prohibited items is slightly higher than that of other age groups, but children and elderly passengers require special care (such as the risk of being unaccompanied). Therefore, the score setting reflects differentiation rather than simply "low risk".

[0106] Existing statistical data does not show a significant difference between gender and security screening risk, so they are assigned the same score and are only used as a supplementary feature for identity integrity.

[0107] Psychological research indicates that abnormal emotions (such as tension and fear) may be related to the intention to carry contraband or evade inspection. The scoring system reflects a positive correlation between the degree of emotional abnormality and risk.

[0108] If a passenger's movement path in the pre-security check area deviates from the standard route (such as bypassing security equipment or lingering), it may indicate an attempt to evade inspection, and the score will decrease according to the degree of deviation.

[0109] Excessive speed may indicate an attempt to rush through, while excessively slow speed may delay inspections. Abnormal stops may involve suspicious behavior. The scoring system reflects the penalties for abnormal speed patterns.

[0110] It should be noted that the movement trajectory features, passage speed features, and facial expression analysis features used in the second feature layer in this step are derived from real-time video analysis and AI instant recognition performed by the pre-security check system when passengers pass through the gate. The output results are discrete qualitative labels (such as normal / abnormal trajectory, normal / too fast / too slow speed, natural / nervous expression, etc.). The above features differ from the continuous spatiotemporal trajectory point sequence and facial image sequence used in the subsequent dynamic trajectory calibration step in terms of data source, accuracy, and purpose: the former is used for the rapid calculation of the initial fusion feature values, while the latter is used for secondary fine calibration of users in the first processing channel.

[0111] (III) Third characteristic layer: historical risk characteristics

[0112] This feature layer has a fifteenth preset weight value (e.g., 20%), built based on historical security data, used to identify users with a long-term risk background. Feature design principle explanation: This feature layer includes three features: historical security records, preset identified personnel, and records of disciplinary actions for breach of trust, all sourced from an authorized historical security data platform. Historical security records and preset identified personnel are high-risk identifiers recognized by authoritative departments, possessing the highest risk weight and subject to a veto mechanism; records of disciplinary actions for breach of trust reflect the traveler's social credit status, and while not directly equivalent to a security threat, can serve as a supplementary reference indicator. These features provide authoritative data support for assessing a traveler's long-term risk background.

[0113] The quantification rules and preset scores for each secondary feature are as follows:

[0114] Historical security record characteristics: A sixteenth preset weight value (e.g., 8%) is set. When there is no historical security record, a first score is assigned (e.g., +10 points); when there is a general violation record, a second score is assigned (e.g., -5 points); when there is an abnormal record, a third score is assigned (e.g., -10 points), and an abnormal category determination is directly triggered.

[0115] Preset personnel identification characteristics: There is a seventeenth preset weight value (e.g., 8%). When the personnel are not identified by the preset labels, they are given the first attention score (e.g., +10 points); when they belong to the seven preset personnel categories such as the preset attention type, they are given the second attention score (e.g., -10 points), and the abnormal category judgment is triggered directly.

[0116] Characteristics of Dishonesty Punishment Records: An eighteenth preset weight value (e.g., 4%) is set. A first dishonesty score (e.g., +5 points) is assigned when there is no dishonesty record; a second dishonesty score (e.g., -3 points) is assigned when there is general dishonesty; and a third dishonesty score (e.g., -5 points) is assigned when there is serious dishonesty.

[0117] Scoring logic explanation:

[0118] Abnormal records (especially those listed as pre-selected targets of attention) indicate that the traveler poses a high potential threat and are subject to a veto; general violation records will be penalized with appropriate point deductions.

[0119] The list of personnel with pre-defined identification is determined by relevant authorities based on intelligence. They pose a high potential threat and must be directly transferred to the manual review channel.

[0120] A record of dishonesty reflects a traveler's social credit status. Although it is not directly equivalent to a security threat, it can serve as a supplementary risk indicator. Those with serious dishonesty will have more points deducted.

[0121] (iv) Fourth Feature Layer: Historical Travel Characteristics

[0122] This feature layer has a nineteenth preset weight value (e.g., 10%), which is built based on historical flight record data from the past three years and is used to assess the stability of passengers' travel habits.

[0123] Feature Design Principles Explained: This feature layer includes four features: flight frequency, security violation records, route stability, and ticketing behavior. Flight frequency reflects a passenger's travel habits; high-frequency passengers typically exhibit stable behavior patterns, are familiar with security procedures, and pose a lower risk. Security violation records are direct evidence of past behavior; repeat offenders have a significantly higher probability of committing the same offense again. Route stability and ticketing behavior reflect a passenger's travel purpose; abnormal behavior (such as frequent changes of destination or frequent ticket cancellations and rebookings) may indicate unstable plans or an intention to evade security checks.

[0124] The quantification rules and preset scores for each secondary characteristic are as follows: Flight frequency characteristic: A twentieth preset weight value is set (e.g., 3%). Different preset scores are assigned based on the number of flights per year. For example, ≥12 flights per year are assigned the first frequency score (e.g., +5 points), 6-11 flights are assigned the second frequency score (e.g., +3 points), 1-5 flights are assigned the third frequency score (e.g., +2 points), and the first flight is assigned the fourth frequency score (e.g., +1 point).

[0125] Security check violation record characteristics: A 21st preset weight value (e.g., 3%) is set. No violation record is assigned the first violation score (e.g., +5 points); one minor violation is assigned the second violation score (e.g., -2 points); multiple violations or serious violations are assigned the third violation score (e.g., -5 points).

[0126] Route stability characteristics: A twenty-second preset weight value (e.g., 2%) is set. When flying a fixed route frequently, a first stability score (e.g., +3 points) is assigned; when the route changes frequently, a second stability score (e.g., +1 point) is assigned; and when frequently changing destinations, a third stability score (e.g., -2 points) is assigned.

[0127] Ticket purchase behavior characteristics: A 23rd preset weight value (e.g., 2%) is set. When the ticket purchase time is more than the first preset time threshold (e.g., 24 hours) before the flight departure, and there is no record of refund or rescheduling, it is judged as a normal ticket purchase and assigned a first ticket purchase score (e.g., +2 points); when the ticket purchase time is less than or equal to the first preset time threshold (e.g., 24 hours) before the flight departure, it is judged as a temporary ticket purchase and assigned a second ticket purchase score (e.g., +1 point); when the total number of refund and rescheduling operations exceeds the preset number threshold (e.g., 3 times) within a preset time window (e.g., the most recent 30 days), it is judged as frequent refund and rescheduling and assigned a third ticket purchase score (e.g., -2 points).

[0128] Scoring logic explanation:

[0129] Frequent passengers are usually familiar with security check procedures and have stable behavior patterns, so the risk is relatively low; first-time passengers may exhibit abnormal behavior (not malicious) due to unfamiliarity with the procedures, but there is a lack of historical data to support this, so their scores are slightly lower than those of frequent passengers.

[0130] Violation records are direct evidence of past behavior. Repeat offenders are significantly more likely to repeat their offenses, and the penalty points increase with the severity of the violation.

[0131] Stable flight routes typically correspond to business or commuter travelers whose behavior is predictable; frequent changes of destination may involve unconventional travel purposes (such as tourism, visiting relatives, etc.), with a neutral to low risk; no excessive penalty points.

[0132] Last-minute ticket purchases and frequent ticket changes may be related to unstable plans or evasion of checks, but they are not absolutely abnormal and only result in minor deductions.

[0133] (v) Fifth Feature Layer: Time and Flight Attribute Features

[0134] This feature layer has a 24th preset weight value (e.g., 5%), which is built based on the attribute data of the current flight and is used to reflect the impact of the spatiotemporal environment on risk.

[0135] Feature Design Principles Explanation: This feature layer includes four features: flight time, sensitive routes, holiday factors, and temporary flights. Flight time affects the allocation of security personnel and passenger behavior patterns (e.g., late-night passengers may behave abnormally due to fatigue); specific route areas are listed as pre-defined control targets due to geographical or special location factors; security levels are increased during special security control periods, increasing the probability of threats; passenger information collection for temporary or charter flights is incomplete, resulting in information uncertainty.

[0136] The quantification rules and preset scores for each secondary feature are as follows:

[0137] Flight time period characteristics: There is a 25th preset weight value (e.g., 2%). Daytime (e.g., 6:00-18:00) is assigned a first time period score (e.g., +3 points), early morning (e.g., 0:00-6:00) is assigned a second time period score (e.g., +2 points), and late night (e.g., 18:00-24:00) is assigned a third time period score (e.g., +2 points).

[0138] Sensitive route characteristics: A 26th preset weight value (e.g., 1%) is set. Domestic regular routes are assigned a first route score (e.g., +2 points), international routes are assigned a second route score (e.g., +1 point), and specific route areas are assigned a third route score (e.g., -2 points).

[0139] Holiday Factor Characteristics: A 27th preset weight value (e.g., 0.5%) is set. Non-holidays are assigned a false score in the first section (e.g., +1 point), holidays are assigned a false score in the second section (e.g., 0 points), and special security control periods are assigned a false score in the third section (e.g., -1 point).

[0140] Temporary flight characteristics: A 28th preset weight value is set (e.g., 0.5%). Regular flights are assigned a first temporary score (e.g., +1 point), while temporary or charter flights are assigned a second temporary score (e.g., -1 point).

[0141] Scoring logic explanation:

[0142] During the daytime, there are sufficient security personnel and comprehensive monitoring, and most passengers are traveling normally. Passengers in the early morning and late at night may exhibit abnormal behavior due to fatigue or rushing to catch a red-eye flight, but the risk difference is not significant, so the score is slightly lower than that during the day.

[0143] Certain routes and regions have been designated as pre-controlled areas due to their special geographical location, and corresponding risk bonuses (deductions) are applied accordingly; international routes are slightly riskier than regular domestic routes due to their involvement in immigration management.

[0144] During periods of special security control, the security level is raised, increasing the probability of potentially high-risk individuals among passengers, thus a risk deduction is set; during holidays, although passenger flow is large, the threat has not increased significantly, so it is set to neutral.

[0145] Passenger information collection for temporary / charter flights is often incomplete and may pose a risk of group-based abnormal behavior; therefore, a slight deduction is set to reflect the uncertainty of the information.

[0146] (vi) Calculation of fused eigenvalues

[0147] The fusion feature value equals a preset base score (e.g., 50 points) plus the sum of the weighted scores of each secondary feature. The sum of weighted scores is the cumulative sum of the weighted scores of all secondary features. The weighted score of each secondary feature is obtained by multiplying its score by its corresponding preset weight value. Each preset weight value is expressed as a percentage and converted to a decimal during calculation (e.g., 6% is converted to 0.06). The scores of each secondary feature are determined according to the aforementioned quantification rules. The purpose of setting the preset base score is to ensure that all target users still have a non-negative initial feature value even without receiving any positive bonus points, avoiding scoring anomalies due to missing feature data.

[0148] (vii) Quick Exit Mechanism

[0149] Adopt a tiered fast exit mechanism:

[0150] Level 1 Exit: When calculating the first feature layer, if an anomaly is triggered by explosion-proof detection, combustible volatile substance detection, radioactive material detection, or controlled contraband detection, the target user is immediately determined to be in the abnormal category, the calculation of all subsequent feature layers is stopped, and the user is directly assigned to the fourth processing channel.

[0151] Second-level exit: When calculating the third feature layer, if the historical security record or the preset identification personnel feature triggers an anomaly, the target user is immediately determined to be in the anomaly category, subsequent calculations are stopped, and the user is directly assigned to the fourth processing channel.

[0152] Level 3 Exit: During the calculation process, if the accumulated weighted deduction score exceeds the preset deduction threshold (e.g., 30 points), it will be automatically identified as an abnormal category, and there is no need to complete the calculation of all feature layers.

[0153] Through the aforementioned fast exit mechanism, most (e.g., 90%) of the target users can complete the calculation of fused feature values ​​within the first preset time threshold (e.g., 100 milliseconds), and the determination time of abnormal category users can be controlled within the second preset time threshold (e.g., 50 milliseconds).

[0154] (viii) Dynamic weight adjustment mechanism

[0155] The multi-level weighted calculation model also supports dynamic weight adjustment:

[0156] Time period adjustment: During the preset morning peak hours (e.g., 7:00-9:00), the preset weight value of the first feature layer is increased to the first weight adjustment value (e.g., 45%), and the preset weight value of the fourth feature layer is reduced accordingly; During the preset late night hours (e.g., 23:00 to 6:00 the next day), the preset weight value of the third feature layer is increased to the second weight adjustment value (e.g., 25%).

[0157] Special event adjustments: When the airport enters a preset special security control state (the special security control state includes, but is not limited to: the Level 1 security response period determined by a specific competent authority, the period of holding a large-scale international conference or event, the period of an important celebration, and other special periods announced in advance by the competent authority), the preset weight values ​​of all security-related feature layers will be increased by the preset ratio (e.g., 10% to 20%). During preset holidays, the preset weight value of the fourth feature layer will be reduced and the preset weight value of the first feature layer will be increased.

[0158] Learning optimization: The preset weight values ​​of each feature layer are automatically adjusted based on the actual security inspection feedback results, and the weight configuration is updated by means of a multi-strategy parallel comparison mechanism (i.e., multiple candidate weight configurations are applied at the same time, and their classification accuracy and false alarm rate are evaluated respectively within the same time window, and the configuration with the best effect is selected as the update parameter).

[0159] Through the weighted calculation of the above five feature layers and the fast exit mechanism, the fusion feature value and anomaly category identifier of the target user are obtained, providing a quantitative basis for subsequent channel allocation and dynamic trajectory calibration.

[0160] S400, based on the fusion feature value, determine the initial processing channel level corresponding to the target user, the initial processing channel level including a first processing channel level corresponding to the fusion feature value being greater than or equal to a first threshold.

[0161] In this application, the processing channel level refers to the security check category assigned to the target user, while the processing channel refers to the actual physical or logical channel resources within the airport security check area corresponding to different levels. Different processing channel levels correspond to different security check procedures and security check intensities, specifically achieved by guiding the target user to the corresponding processing channel.

[0162] In this application, the security inspection intensity is used to comprehensively characterize the technical configuration of the security inspection process, including the following quantifiable dimensions:

[0163] Human intervention rate: This refers to the proportion of inspection operations performed manually out of the total number of inspection operations during the security check process. The first processing channel is mainly automated equipment, with the lowest human intervention rate; the fourth processing channel is mainly one-on-one manual inspection, with the highest human intervention rate.

[0164] Inspection execution depth: refers to the level of detail in the inspection of passengers and their luggage, such as whether to require opening luggage, whether to require removing shoes, and whether to conduct a full-body hand search.

[0165] Inspection time: refers to the average time required to complete the entire security check process. In this application, the processing channels are divided into four levels: Level 1 (corresponding to the fast-track channel), Level 2 (corresponding to the standard security check channel), Level 3 (corresponding to the enhanced security check channel), and Level 4 (corresponding to the manual review channel). The security inspection intensity corresponding to the Level 1, Level 2, Level 3, and Level 4 increases progressively. The security configuration and processing rules corresponding to each channel level are as follows.

[0166] (a) Rules for determining the initial processing channel level

[0167] Based on the fusion feature value calculated in step S300 and whether the preset abnormal feature (i.e., the abnormal category identifier) ​​is triggered, the initial processing channel level of the target user is determined according to the following rules:

[0168] 1. First-level processing lane (fast track) admission criteria:

[0169] When the fused feature value is greater than or equal to the first preset feature value threshold (e.g., 85 points), and no preset abnormal features are triggered (i.e., no positive results for explosion-proof detection, controlled prohibited items, abnormal records, preset identified personnel, etc.), and the identity verification is passed, and the security check violation records in the historical travel features are no violations or only minor violations, it is determined to be at the first processing channel level. This channel corresponds to fast-pass security check, with the lowest security intensity and the highest passage efficiency.

[0170] 2. Admission requirements for the second processing lane level (standard security checkpoint):

[0171] When the fused feature value is less than the first preset feature value threshold and greater than or equal to the second preset feature value threshold (e.g., 70-84 points), and no major risk factors are triggered (e.g., no abnormal records, no preset indicators), minor abnormalities are allowed (e.g., carrying general prohibited items, single minor violation records), and the channel is classified as the second processing channel level. This channel follows the standard civil aviation security check procedure, without simplification or upgrade.

[0172] 3. Admission requirements for the third-level processing lane (enhanced security checkpoint):

[0173] If the fusion feature value is less than the second preset feature value threshold and greater than or equal to the third preset feature value threshold (e.g., 55-69 points), or if one or more medium-risk factors exist (such as multiple security check violations, highly sensitive flight routes, frequent ticket refunds and changes), it is determined to be at the third processing channel level. This channel adopts a comprehensive inspection process, adding manual inspection and baggage opening inspection.

[0174] 4. Admission criteria for the fourth processing channel level (manual review channel):

[0175] When the fusion characteristic value is less than the third preset characteristic value threshold (e.g., below 55 points), or any preset abnormal characteristic is triggered (including positive explosion detection, controlled contraband, abnormal records, preset identified personnel, positive radioactive material, etc.), or the pre-security inspection system detects a major abnormality (such as a suspected explosive simulation), it is directly determined to be at the fourth processing channel level. This channel adopts strict one-on-one manual inspection, and if necessary, it will be transferred to the relevant competent authorities for handling.

[0176] In this embodiment, a minor violation refers to a passenger carrying general prohibited items due to negligence or without malicious intent, and who actively cooperates with the security personnel after being alerted, without causing interruption of the security check process or significant disruption to order. A general violation refers to a passenger intentionally concealing prohibited items, carrying a large quantity of prohibited items, or causing some impact on security check order (such as non-cooperation with inspection, or a bad attitude). A serious violation refers to a passenger carrying controlled prohibited items (including but not limited to controlled knives, stun guns, replica guns, explosives, flammable and explosive materials, radioactive materials, etc.), or forcibly rushing through the security checkpoint. Multiple violations refer to two or more violations (regardless of level) within a preset time window (e.g., the most recent 12 months). General prohibited items refer to items that are not classified as controlled but are prohibited from being carried on board by civil aviation security regulations, including but not limited to lighters, matches, liquids exceeding 100ml per unit, ordinary fruit knives (blade length less than 6cm), and blunt objects (such as small hammers).

[0177] (II) Rules for handling special passenger groups

[0178] For children (e.g., 12 years and under) and older adults (e.g., 65 years and over), add peer group determination logic:

[0179] Grouping determination: If the target user and one or more other passengers meet any of the following conditions, they are determined to be in the same group:

[0180] a) The flight information is the same and the reservation record number (PNR) is the same;

[0181] b) Although they do not have the same PNR, their passage times are similar (e.g., the time difference between passing through the pre-security gate is less than the preset time window, such as 30 seconds) and they are identified by the system as being in the same line.

[0182] Unaccompanied minors: Children or elderly people without adult accompaniment will still be assigned to the normal channel level determined by their own integration characteristics, or guided to the preset special care channel.

[0183] Unified channel for peer groups: Within the same peer group, compare the risk level of each member (i.e., the channel level corresponding to the fusion feature value), and select the member with the highest level as the basis for the allocation of the whole group; if any member in the group triggers an abnormal category (one-vote veto high risk), the whole group will be directly assigned to the fourth processing channel (manual review channel), and the scores will no longer be compared separately.

[0184] (III) Security check configuration of each processing channel

[0185] Based on the determined channel level, the target user is guided to the corresponding physical channel and the appropriate security check procedure is performed:

[0186] First processing channel (fast passage channel) configuration:

[0187] Identity verification: Just scan your face and you're done; no secondary manual verification is required.

[0188] Personal security check: No need to remove coat or belt, use high-speed millimeter wave gate for seamless passage, no manual inspection is required if there is no alarm.

[0189] Baggage security check: Power banks must be removed separately for inspection; items such as laptops and umbrellas do not need to be removed and can be passed directly through the X-ray machine or CT scanner; when the baggage X-ray machine image interpretation system does not output an abnormal alarm, the system will automatically release the baggage without opening it or manual re-inspection.

[0190] Traffic flow: Prioritize passage, reduce queuing and waiting, perform only standard equipment detection, and do not increase human intervention.

[0191] Configuration of the second processing lane (standard security check lane):

[0192] Identity verification: machine verification plus simple visual review by security personnel.

[0193] Personal Security Check: Pass through the security gate or millimeter-wave gate normally; if the equipment alarms, a handheld detector will be used for targeted manual inspection and key touch checks; if there is no alarm, no manual inspection will be conducted. Baggage Security Check: Items requiring separate inspection, such as laptops, power banks, liquids, and umbrellas, must be removed and inspected through the X-ray machine; if the X-ray machine does not detect any abnormalities, the baggage will not be opened or manually opened. General Principle: Follow standard civil aviation security checks; neither simplify nor upgrade them.

[0194] The third processing lane (enhanced security check lane) is configured as follows:

[0195] Identity verification: Machine verification plus strict manual comparison of identity documents and faces.

[0196] Security check: A thorough hand check of all personnel, with a focus on the armpits, waist, ankles, and inner layers of clothing.

[0197] Baggage security check: All baggage must undergo a thorough re-inspection, and any suspected items must be opened and manually inspected; personal belongings, backpacks, and handbags must be emptied and inspected, and liquids must be opened and verified one by one.

[0198] General principle: Simplify and restrict release, do not enjoy any exemption or removal policies, and conduct manual verification item by item.

[0199] Configuration of the fourth processing channel (manual review channel):

[0200] Guidance and Isolation: Passengers will be guided to a separate isolation security checkpoint, physically separated from ordinary passengers.

[0201] Security checks: One-on-one manual full-body checks will be conducted to thoroughly inspect any concealed items.

[0202] Baggage security check: 100% of baggage is opened and thoroughly inspected, and all items are checked one by one.

[0203] Information verification: Simultaneously verify identity information and flight information, and register and retain security check records.

[0204] Handling and Transfer: When there are major hidden dangers such as pre-defined labeling type or positive dangerous goods test, the goods shall be directly transferred to the relevant departments for handling, and it is forbidden to release them on one's own.

[0205] (iv) Dynamic scheduling mechanism for channels

[0206] The system monitors the operational status of each processing channel in real time and dynamically adjusts channel resource allocation to adapt to changes in passenger flow.

[0207] Channel capacity management: Configure the number of channels of each type according to the preset capacity ratio. For example: the first processing channel (fast) accounts for 30%-40% of the total number of channels; the second processing channel (standard) accounts for 40%-50% of the total number of channels; the third processing channel (enhanced) accounts for 10%-15% of the total number of channels; and the fourth processing channel (manual review) accounts for 5%-10% of the total number of channels. The actual ratio can be dynamically adjusted according to airport operation data.

[0208] Adjust strategies in real time:

[0209] When the queue length of the first processing channel exceeds the first preset congestion threshold (e.g., the number of people in the queue exceeds 80% of the channel capacity), or the passage waiting time exceeds the preset time threshold (e.g., 5 minutes), some of the second processing channels will be automatically temporarily configured as the first processing channel, and the admission rules will be adjusted simultaneously (e.g., temporarily lowering the fusion feature value threshold, but still retaining the bottom line condition of no high-risk triggers).

[0210] A severe congestion state is identified when the queue length in any processing lane exceeds a second preset congestion threshold (e.g., 120% of lane capacity), the average waiting time exceeds a second preset time threshold (e.g., 10 minutes), or the passenger density in the lane exceeds a preset density threshold (e.g., 4 people / square meter). At this time, a diversion mechanism is automatically activated, reassigning subsequent target users to other available lanes that meet the security check intensity requirements. Once the peak passenger flow period ends, the original lane configuration is automatically restored.

[0211] Through the aforementioned channel allocation and scheduling mechanism, security check resources are dynamically matched with passenger risk levels, thereby improving overall traffic efficiency while ensuring safety.

[0212] S500, when the initial processing channel level is the first processing channel level, the initial processing channel level is adjusted based on the spatiotemporal trajectory data of the target user, and the adjusted processing channel level is used as the classification result of the target user.

[0213] In this embodiment, when the initial processing channel level is not the first processing channel level, the initial processing channel level is directly used as the classification result of the target user, and dynamic trajectory calibration is not performed. In this embodiment, dynamic trajectory calibration is only performed on the first processing channel level because this level corresponds to low-risk users, and after granting fast access permissions, secondary security verification still needs to be performed through high-precision trajectory behavior analysis; for other levels, the security check intensity is already high, and no additional calibration is required.

[0214] Specifically, adjusting the initial processing channel level based on the spatiotemporal trajectory data of the target user includes:

[0215] S501, the spatiotemporal trajectory data is divided into multiple trajectory segments, and the time information corresponding to each trajectory segment is recorded.

[0216] In this step, the division of the trajectory segment can be achieved using one or a combination of the following methods.

[0217] Method 1: Uniform Time Division

[0218] In the uniform time division strategy, the spatiotemporal trajectory data of the target user is divided into multiple trajectory segments of equal duration according to a fixed time length.

[0219] Specifically, a preset segment duration parameter L (e.g., 60 seconds) is used. Starting from the trajectory start time, consecutive subsequences of duration L are taken as trajectory segments until the entire target time interval is covered. The duration of the last trajectory segment may be less than L, and it will be processed according to the actual duration.

[0220] Each trajectory segment records the following information: segment number, start timestamp, end timestamp, time interval from the flight departure time (taking the remaining time at the midpoint of the segment or the end of the segment), and sequence of trajectory points within the segment.

[0221] The uniform time partitioning strategy is simple to implement and has low computational complexity, making it suitable for scenarios where there is no prior knowledge of passenger behavior patterns or where large amounts of data need to be processed quickly.

[0222] Method 2: Non-uniform time division

[0223] In the non-uniform time division strategy, different trajectory segments have different durations, and the longer the time interval between the trajectory segment and the expected event time (i.e., the flight departure time), the longer the duration of the trajectory segment; the closer the trajectory segment is to the expected event time, the shorter its duration.

[0224] The specific division method is as follows:

[0225] Let the flight departure time be T. takeoff The trajectory acquisition deadline is T. end =T takeoff -ΔT (where ΔT is a preset time threshold, such as 20 minutes). The start time for the target user's pre-security check to be completed is T. start From T start To T end Within the time window, segment according to the remaining duration:

[0226] When the remaining time exceeds the first preset time threshold (e.g., 30 minutes), the first preset segment duration (e.g., 300 seconds) is used.

[0227] When the remaining time is between the first preset time threshold and the second preset time threshold (e.g., 15–30 minutes), the second preset time segment (e.g., 120 seconds) is used.

[0228] When the remaining time is less than the second preset time threshold (e.g., 15 minutes), the third preset time segment (e.g., 30 seconds) is used.

[0229] Among them, the duration of the first preset segment is greater than the duration of the second preset segment, which is greater than the duration of the third preset segment.

[0230] The technical purpose of the non-uniform partitioning strategy is as follows: In the phase far from the departure time, passenger behavior is more random and there is ample time. Using longer trajectory segments can extract stable behavior patterns and reduce computational overhead. In the phase closer to the departure time, passengers may accelerate their behavior or exhibit abnormalities due to rushing to catch their flights. Using shorter trajectory segments can improve the sensitivity and timeliness of anomaly detection.

[0231] In another illustrative embodiment, uniform and non-uniform partitioning strategies can be combined. Specifically, the basic trajectory segments are first divided according to the non-uniform strategy, and then the interior of each basic trajectory segment is uniformly subdivided according to a fixed time resolution to balance behavioral stability and local anomaly detection capability.

[0232] Method 3: Division of Uniform and Non-uniform Mixtures

[0233] In another illustrative embodiment, uniform and non-uniform partitioning strategies can be combined. Specifically, the basic trajectory segments are first partitioned using a non-uniform strategy, and then the interior of each basic trajectory segment is uniformly subdivided at a fixed time resolution to balance behavioral stability and the ability to detect local anomalies.

[0234] This hybrid strategy retains the adaptability of non-uniform segmentation to the overall behavior pattern (the longer the segment is further from takeoff, the shorter the segment is closer to takeoff), while ensuring the sampling consistency of local behavior through internal uniform sub-segments, which facilitates subsequent feature extraction and model comparison.

[0235] Method 4: Event Trigger Division

[0236] In another illustrative embodiment, in addition to dividing the time into segments based on a fixed duration, the segments can also be triggered by behavioral events. For example:

[0237] Area Entry Trigger: When a target user enters a preset key area (such as a security check queue area, document verification area, or baggage claim area), the start of a new trajectory segment is automatically triggered. The system determines whether the user has entered the key area by matching their location coordinates with the preset area boundaries.

[0238] Behavioral change trigger: When the target user's behavioral characteristics change significantly, the trajectory segment is automatically cut at the point of change. For example:

[0239] The dwell time exceeds the third preset time threshold (e.g., staying stationary in a non-queue area for more than 30 seconds).

[0240] The movement speed suddenly changes beyond the preset speed change rate (for example, the speed drops suddenly from 1 m / s to 0.2 m / s within 5 seconds, or increases suddenly from 0.5 m / s to 2 m / s).

[0241] The direction angle changes drastically (such as a sudden reversal or a large-angle turn).

[0242] Event-triggered segmentation can flexibly capture key changes in passenger behavior, and is particularly suitable for identifying abnormal behaviors (such as sudden stops, acceleration, and reversals), improving the accuracy of subsequent trajectory pattern comparisons. Regardless of the segmentation strategy used, the maximum number of trajectory segments is limited by a preset maximum number of segments (e.g., 20 segments). If the number of segments divided according to the above rules exceeds the maximum number of segments, adjacent trajectory segments furthest from the departure time are merged in chronological order until the maximum number of segments is met; if the number of segments is insufficient, the existing segmentation is maintained.

[0243] Each trajectory segment records the following information: segment number (incrementing from 1); start and end timestamps; the time interval from takeoff for this trajectory segment (the remaining time at the midpoint of the segment or the end of the segment); and the sequence of trajectory points within this trajectory segment (including the spatial coordinates and timestamps of each point).

[0244] S502, extract features from each trajectory segment to generate the corresponding trajectory feature vector.

[0245] For each trajectory segment obtained in step S501, numerical features that characterize the behavior pattern of the trajectory segment are extracted, and multiple features are combined into a trajectory feature vector of fixed dimensions. Feature extraction covers multiple dimensions such as geometric shape, motion attributes, statistical distribution, and behavioral semantics. The specific calculation of feature extraction adopts conventional numerical calculation methods in this field (such as Euclidean distance to calculate path length, arithmetic mean to calculate center point, etc.), without relying on specific dedicated hardware or non-publicly known algorithms.

[0246] (a) Geometric morphological characteristics

[0247] Based on the spatial coordinate sequence of each trajectory point within the trajectory segment, geometric features describing the shape of the path are extracted:

[0248] Total path length: the sum of Euclidean distances between adjacent trajectory points within a trajectory segment.

[0249] Displacement distance: The straight-line distance between the start and end points of a trajectory segment.

[0250] Tortuousness: The ratio of total path length to displacement distance, reflecting the degree of curvature of the trajectory.

[0251] Change in direction angle: The sum of the absolute values ​​of the cumulative change in direction angle within the trajectory segment.

[0252] Bounding box area: The area of ​​the smallest bounding rectangle of the trajectory point sequence.

[0253] (ii) Motion Attribute Characteristics

[0254] Based on the spatial coordinates and timestamps of trajectory points, features describing the movement dynamics are extracted:

[0255] Average speed: Total path length divided by total duration of track segment.

[0256] Velocity variance: The variance of velocity at each instant, reflecting the degree of velocity fluctuation.

[0257] Maximum instantaneous velocity: The maximum moving velocity between adjacent trajectory points.

[0258] Average acceleration: change in velocity divided by time.

[0259] Dwell percentage: The proportion of trajectory points with speeds below a preset stationary threshold (e.g., 0.1 m / s).

[0260] (III) Spatiotemporal distribution characteristics

[0261] Based on the distribution of trajectory points in the spatiotemporal dimension, statistical features are extracted:

[0262] Duration: The end timestamp of the trajectory segment minus the start timestamp.

[0263] Number of trajectory points: The number of trajectory points contained in this trajectory segment.

[0264] Spatial center point: the average value of the coordinates of the trajectory points.

[0265] Spatial divergence: The average distance between a point on the trajectory and the center point.

[0266] Directional consistency: The circular variance of the direction angle of the trajectory points; the smaller the value, the more consistent the movement direction.

[0267] (iv) Behavioral semantic features

[0268] Extracting high-level features with behavioral discrimination significance from trajectory segments:

[0269] Loitering Index: The number of times a trajectory point repeatedly visits the same area (referring to a circular area with a preset radius (e.g., 2 meters) centered on a trajectory point, or a preset grid cell). Extraction Method:

[0270] Divide the two-dimensional space where the target user is located into grid cells of a preset size (e.g., a square grid with a side length of 2 meters), or perform adaptive clustering based on the distribution of trajectory points.

[0271] Traverse each trajectory point within the trajectory segment to determine its corresponding grid cell or cluster.

[0272] Count the number of times the same grid cell or cluster is visited (multiple trajectory points that remain continuously in the same cell are counted as one visit).

[0273] The number of visits to all units within the trajectory segment is sorted or summed to obtain the wandering index.

[0274] A feasible alternative is to calculate the minimum distance between a trajectory point and its historical trajectory points, and count the number of times the minimum distance is less than a preset threshold (e.g., 1 meter).

[0275] Duration of stay in designated areas: The cumulative time spent in preset key areas (such as security check queues and document verification areas). Extraction method:

[0276] Predefine the spatial scope of key areas, such as security check queuing areas, document verification areas, and restroom entrances, and store them in the form of polygonal or circular bounding boxes.

[0277] Iterate through each trajectory point within the trajectory segment and determine whether the coordinates of the trajectory point fall within the bounding box of any key region.

[0278] The durations corresponding to the sequence of trajectory points that fall consecutively within the same key area are accumulated (the time difference between the first and last points).

[0279] Output the cumulative stay time in each key area, or output the total stay time in all key areas.

[0280] Abnormal status indicators: whether there is a preset abnormal behavior pattern, such as sudden reversal, frequent entry and exit from a certain area, or sudden acceleration after a long period of stillness.

[0281] The method for determining a sudden reversal is as follows:

[0282] Calculate the change in the direction angle of movement of adjacent trajectory points.

[0283] If the change in direction angle exceeds a preset angle threshold (e.g., 150 degrees), and the displacement distances corresponding to both directions exceed a preset distance threshold (e.g., 1 meter), it is determined to be a sudden reversal.

[0284] If at least one sudden reversal occurs within the trajectory segment, the corresponding bit or label is set in the abnormal status flag.

[0285] Frequent entry and exit from a certain area refers to the number of times a target user enters and leaves a preset key area (such as the entrance to a restroom, a shopping area, or the edge of a security checkpoint queue) within a trajectory segment, exceeding a preset threshold (e.g., 3 times). Entry is determined by the trajectory point coordinates entering the bounding box of the area, and exit is determined by the coordinates moving out of the bounding box.

[0286] Sudden acceleration after a long period of stillness refers to a situation where the target user's moving speed is continuously lower than a preset still speed threshold (e.g., 0.1 m / s) for a period of time, and the stillness duration exceeds a preset stillness duration threshold (e.g., 10 seconds), and then its instantaneous speed jumps to a preset acceleration speed threshold (e.g., 1.5 m / s) or higher within a preset short window (e.g., 2 seconds).

[0287] It should be noted that the above methods for extracting behavioral semantic features are all based on well-known principles of spatial querying, rule determination, and statistical counting, and do not rely on non-known algorithms or specialized hardware. For example, determining whether a point is within a polygonal region is a classic problem in computational geometry (ray casting or winding number method), mesh generation and clustering are conventional data preprocessing methods, and rule determination is a basic logical operation. Those skilled in the art can extract the corresponding features by following the above guidelines without any creative effort.

[0288] (v) Feature vector generation

[0289] All the extracted numerical features are concatenated into a fixed-dimensional feature vector in a preset order, and a trajectory feature vector is generated for each trajectory segment. Features that cannot be calculated are filled with preset default values.

[0290] Using the feature extraction method described above, each trajectory segment is converted into a numerical trajectory feature vector, providing input for subsequent clustering model comparison.

[0291] S503, the trajectory feature vector of each trajectory segment is input into a clustering model associated with the time information of that trajectory segment. The output of the clustering model includes at least one first-class cluster result and one second-class cluster result.

[0292] Different clustering models correspond to different time information ranges (i.e., different time intervals from the expected event time), and the clustering radius threshold of the clustering model corresponding to the trajectory segment with a shorter time interval from the expected event is smaller.

[0293] (I) Construction and Training of Clustering Models

[0294] For different time information ranges, corresponding clustering models are pre-trained. Each clustering model is built independently to adapt to the differences in passenger behavior patterns at different stages.

[0295] 1. Division of Time Information Range

[0296] The time window from the completion of pre-security check to entering the waiting area (or until a preset threshold before takeoff) for the target user is divided into several consecutive time intervals. For example:

[0297] Section 1: More than 30 minutes until departure;

[0298] Section 2: Between 15 and 30 minutes before departure;

[0299] Section 3: Less than 15 minutes before takeoff.

[0300] Each time interval corresponds to an independent clustering model.

[0301] 2. Collection of training data

[0302] For each time interval, a large number of historical trajectory segments of normal passengers (i.e., users with no violation records and not identified as high-risk) are collected within that time interval, and trajectory feature vectors are extracted according to the S502 method to form a training sample set. The training sample set only contains normal behavior samples and does not contain abnormal behavior samples.

[0303] 3. Training the clustering model

[0304] An unsupervised clustering algorithm (such as K-means) is used to cluster the training sample set for each time interval. The input of the clustering algorithm is the trajectory feature vector, and the output is several clusters. Each cluster represents a typical normal trajectory pattern.

[0305] The training process includes:

[0306] Determine the number of clusters K (e.g., K=10). This can be determined automatically using methods such as the elbow rule or profile coefficient, or it can be preset based on experience.

[0307] The K-means algorithm is executed to obtain K cluster centers C1, C2, ..., CK.

[0308] For each cluster j (j takes values ​​from 1 to K), calculate its cluster radius R. j The cluster radius is defined as the maximum distance from all sample points within the cluster to the cluster center, or a preset multiple of the standard deviation of the distance (e.g., 2 times the standard deviation).

[0309] 4. Correlation between cluster radius threshold and time information

[0310] Based on the characteristics of the corresponding time intervals, the cluster radius threshold is scaled. For intervals with shorter time intervals from the expected event, passenger behavior patterns are more restricted (e.g., rushing to catch a flight, direct routes), and the degree of variation in normal behavior is smaller; therefore, a smaller cluster radius threshold is used. For intervals with longer time intervals from the expected event, passenger behavior is more random, and the degree of variation in normal behavior is larger; therefore, a larger cluster radius threshold is used.

[0311] The specific implementation method is as follows: First, calculate the basic clustering radius R based on the training sample set. j base Then multiply by a scaling factor α(t) associated with that time interval. The scaling factor is a decreasing function of the time remaining until takeoff: the longer the interval, the larger α(t); the shorter the interval, the smaller α(t). For example, α(t) = 0.5 + 0.5 × t / T maxWhere t represents the time interval (in minutes) between the end time of the current trajectory segment and the flight's departure, T max This represents the maximum time interval between the completion of pre-security check and the preset threshold time before takeoff. Alternatively, different cluster radius thresholds can be set independently for each time interval.

[0312] (ii) Definition of the output results of the clustering model

[0313] For an input trajectory feature vector, the output of the clustering model is divided into two categories:

[0314] First type of cluster result: When the distance between the input feature vector and a normal cluster is less than the cluster radius of that cluster, the identifier of that cluster (e.g., cluster number 1 to K) is output, indicating that the behavior pattern of the trajectory segment belongs to the corresponding normal trajectory type.

[0315] The second type of clustering result: When the distance between the input feature vector and all normal clusters is greater than or equal to the cluster radius of the corresponding cluster, a preset special identifier (e.g., 0 or -1) is output, indicating that the trajectory segment cannot be classified into any known normal trajectory pattern, that is, it is judged as abnormal.

[0316] (III) Model Call and Output Records

[0317] During runtime, for each trajectory segment obtained in step S501, the time information range to which it belongs is first determined based on the time information recorded in the trajectory segment (e.g., the time interval between the midpoint or end of the segment and takeoff), and then the corresponding clustering model is selected. The trajectory feature vector of the trajectory segment is input into the selected model to obtain the output result.

[0318] The output results, along with the segment number and time information of the trajectory segment, are recorded as input for the subsequent step S504 (determining the trajectory anomaly level).

[0319] Through the above design, each clustering model can perform accurate pattern comparison for behavioral characteristics at a specific time stage, and the clustering radius decreases as the time to takeoff shortens, thereby improving the sensitivity of anomaly detection in the near takeoff stage.

[0320] S504, when the output result is the second type of cluster result, calculate the deviation degree of the trajectory feature vector of the trajectory segment from the first type of trajectory cluster closest to the trajectory feature vector in the feature space, and determine the anomaly level of the trajectory segment based on the deviation degree.

[0321] In this embodiment, the degree of deviation is defined as the ratio of the feature distance between the trajectory feature vector of the trajectory segment and its nearest first-class trajectory cluster to the cluster radius of the nearest first-class trajectory cluster.

[0322] Based on the preset numerical range of the deviation, determine the anomaly level of the trajectory segment:

[0323] First anomaly level: Triggered when the deviation is less than the first preset threshold, indicating that the behavior pattern of the trajectory segment deviates from the normal range to a low degree, and the current processing channel level remains unchanged.

[0324] Second anomaly level: Triggered when the deviation is greater than or equal to the first preset threshold and less than the second preset threshold, indicating a moderate degree of deviation, the processing channel level will be adjusted to the second processing channel level (standard security check channel).

[0325] Third anomaly level: Triggered when the deviation is greater than or equal to the second preset threshold and less than the third preset threshold, indicating a large degree of deviation, the processing channel level will be adjusted to the third processing channel level (enhanced security check channel).

[0326] Fourth anomaly level: Triggered when the deviation is greater than or equal to the third preset threshold, indicating a serious deviation, the processing channel level will be adjusted to the fourth processing channel level (manual review channel).

[0327] Among them, the first preset threshold, the second preset threshold, and the third preset threshold are preset numerical boundaries used to divide different levels of deviation.

[0328] At least one of the first preset threshold, the second preset threshold, and the third preset threshold is a dynamic threshold, and its specific value is dynamically adjusted based on the facial expression recognition results of the target user during the pre-security check phase. The dynamic threshold is determined as follows:

[0329] Facial image acquisition:

[0330] During the pre-security check phase (e.g., when a target user passes through a pre-security check gate), high-definition cameras deployed at or near the gate continuously capture a sequence of facial images of the target user. The frame rate can be 10-30 frames per second, and the capture is continued for a preset duration (e.g., 2-5 seconds) to ensure sufficient dynamic information of facial expressions is obtained.

[0331] Facial feature extraction and classification:

[0332] (1) Extract facial expression feature vectors from the acquired facial image sequence. Feature extraction can be performed using one or a combination of the following methods:

[0333] Geometric features based on facial key points: Several key points on the face are located using facial key point detection algorithms (such as regression tree-based face alignment algorithms or convolutional neural networks), including but not limited to: eyebrows (inner and outer eyebrows), eyes (upper and lower eyelids, inner and outer corners of the eyes), nose (tip and bridge of the nose), lips (corners of the mouth, midpoints of the upper and lower lips), and chin. The two-dimensional coordinates of the key points are extracted, and the following derived features are calculated:

[0334] The distance between key points (such as the distance between eyebrows, the distance between the eyes and mouth);

[0335] The relative angles of key points (such as the angle of eyebrow tilt and the angle of mouth upward);

[0336] The velocity and acceleration of keypoint motion (changes across frames).

[0337] All geometric features are concatenated into a feature vector of fixed dimensions.

[0338] Texture-based deep learning features: Lightweight convolutional neural networks (such as MobileNet and ShuffleNet) are used to perform forward propagation on preprocessed facial images (aligned and normalized), extracting feature maps from intermediate or fully connected layers of the network as deep feature vectors. This network can be pre-trained on large-scale facial expression datasets (such as FER2013 and RAF-DB) to learn general expression representations.

[0339] Facial Action Unit Features: Detect the activation intensity of facial action units (AUs, such as AU1 (inner eyebrow raise), AU2 (outer eyebrow raise), AU4 (eyebrow droop), AU5 (upper eyelid raise), AU6 (cheek raise), AU12 (corner of mouth lift), AU15 (corner of mouth droop), AU23 (lip tightening), etc.). Construct a feature vector (e.g., a 47-dimensional AU intensity vector) by combining specific AU intensity values.

[0340] Regardless of the method used, the final output facial expression feature vector is a fixed-dimensional numerical vector (e.g., 128-dimensional or 256-dimensional), which serves as the input to the facial expression classification model.

[0341] (2) Input the extracted facial feature vectors into the pre-trained facial expression classification model.

[0342] An expression classification model is used to determine the current emotional state category of a target user based on the input expression feature vector. The model structure can adopt any of the following:

[0343] A fully connected neural network consists of an input layer (with dimensions equal to the dimension of the facial expression feature vector), one or more hidden layers (e.g., two layers, each with 64 neurons, using the ReLU activation function), and an output layer. The output layer uses the Softmax activation function to output the probability distribution corresponding to each emotion category.

[0344] Support Vector Machine: Multiple binary classification SVMs are trained using a one-to-many strategy, and the kernel function can be the radial basis function (RBF).

[0345] Random Forest: Integrates multiple decision trees and determines the sentiment category through majority voting.

[0346] Training Process: A large number of face image samples labeled with emotion categories are pre-collected (e.g., from public datasets such as FER2013 and RAF-DB, or self-collected and labeled data). Expression feature vectors are extracted from the samples using the feature extraction method described above. Using the feature vectors as input and the labeled emotion categories as supervision signals, a supervised learning approach is used to train the classification model. After training, the model can output the corresponding emotion category for the input feature vectors, such as "peaceful," "slightly tense," "significantly tense," "anxious," and "fearful."

[0347] Output: The expression classification model outputs a discrete emotion category label, or outputs the probability distribution of each emotion category (the category with the highest probability is selected later).

[0348] Obtaining the correction coefficient: Based on the output emotion state category, the corresponding correction coefficient is obtained from a preset mapping relationship. For example:

[0349] Calm expression: Correction coefficient = 0.8 (reducing the threshold to make the judgment more lenient);

[0350] Slight tension: Correction factor = 1.0 (maintain baseline);

[0351] Significant tension or anxiety: Correction factor = 1.2 (enlarge the threshold to make the judgment more stringent);

[0352] Fear or abnormality: Correction factor = 1.5 (significantly amplifies the threshold).

[0353] Dynamic threshold calculation: The dynamic threshold is obtained by multiplying the preset base threshold (e.g., baseline values ​​1, 2, 3) by the correction coefficient. That is: Dynamic threshold = Base threshold × Correction coefficient.

[0354] In one illustrative embodiment, the first preset threshold is a fixed value and does not participate in dynamic adjustment. It is used to distinguish whether the channel level needs to be upgraded. For example, the first preset threshold is fixed at 1, which means that when the deviation is less than 1, the trajectory feature vector is within the cluster radius of the normal cluster, which is within an acceptable normal fluctuation range, so there is no need to adjust the channel level. The second and third preset thresholds adopt a dynamic adjustment mechanism, and the specific determination method is as follows:

[0355] The second and third basic thresholds are obtained based on the statistical distribution of deviations from historical normal trajectory samples. The specific method is as follows:

[0356] Collect a large amount of trajectory segment data of normal passengers (i.e., historical users with no violation records and not identified as high-risk) within the corresponding time interval, and calculate the deviation degree of each trajectory segment (i.e. the ratio of feature distance to cluster radius) according to steps S502 to S504.

[0357] Perform statistical analysis on the obtained deviations and calculate their quantiles. Determine the preset first quantile (e.g., 95th quantile) as the second basic threshold, and the preset second quantile (e.g., 99th quantile) as the third basic threshold, with the second basic threshold being less than the third basic threshold.

[0358] As an example, the second basic threshold can be 3, and the third basic threshold can be 5. The specific values ​​of the basic thresholds can be pre-calibrated based on the operational data or historical statistical results of different airports, and dynamically updated through incremental learning during system operation. It should be noted that the design logic of the correction coefficient is as follows: when the target user exhibits negative emotions such as tension or anxiety, it indicates that they may be in an abnormal psychological state. At this time, the deviation threshold is appropriately amplified to avoid normal behavioral fluctuations caused by emotions being misjudged as abnormal; conversely, when the expression is calm, the threshold is reduced to increase the sensitivity to abnormal behavior.

[0359] Based on the above steps, the process for determining the anomaly level of a single trajectory segment is as follows:

[0360] Calculate the feature vector of the trajectory segment and the feature distance between it and its nearest normal cluster;

[0361] Calculate the ratio of the feature distance to the cluster radius of the cluster to obtain the degree of deviation;

[0362] The anomaly level of the trajectory segment is determined based on the dynamically adjusted preset threshold range.

[0363] The anomaly level of this trajectory segment will be recorded and used as input for subsequent steps.

[0364] This embodiment achieves the following technical effects by introducing an expression correction threshold mechanism: On the one hand, by dynamically adjusting the judgment threshold for trajectory deviation based on the passenger's real-time emotional state, it can effectively distinguish between normal behavioral fluctuations caused by emotional factors and potential security threats, thus reducing the false alarm rate; on the other hand, this mechanism achieves the fusion of multimodal features (behavioral trajectory + facial expression) with lightweight computational overhead, improving classification accuracy while maintaining real-time processing capabilities, and enhancing the system's adaptability to different scenarios and individual differences.

[0365] S505, by combining the anomaly levels of each trajectory segment, the trajectory anomaly level of the target user is obtained as the comparison result, and the initial processing channel level is adjusted based on the comparison result.

[0366] For all trajectory segments (K segments in total) of the target user, the anomaly level (one of the first to fourth anomaly levels) of each trajectory segment has been determined using S504. The goal of the comprehensive evaluation is to obtain the overall trajectory anomaly level of the user, referred to as the target anomaly level. The comprehensive evaluation can be conducted using one of the following methods:

[0367] The highest anomaly level method: This method selects the highest (i.e., the largest) anomaly level among all trajectory segments as the target anomaly level. For example, if 19 trajectory segments are at the first anomaly level and 1 trajectory segment is at the third anomaly level, then the target anomaly level is the third anomaly level. This method reflects a "strict selection" safety strategy.

[0368] Weighted voting method: Weights are assigned to each trajectory segment, and the target anomaly level is determined based on the weighted voting results. The principle for weight allocation is: the trajectory segment closer to the flight departure time receives a higher weight. Specifically:

[0369] The time window is divided into several intervals. For example, the weight of the trajectory segment more than 15 minutes before takeoff is 0.5, the weight of the trajectory segment 5 to 15 minutes before takeoff is 1.0, and the weight of the trajectory segment within 5 minutes before takeoff is 2.0.

[0370] Calculate the weighted total score for each anomaly level, and take the anomaly level with the highest total score as the target anomaly level.

[0371] If multiple anomaly levels have the same total score, the higher anomaly level will be taken as the target anomaly level.

[0372] It should be noted that both the highest anomaly level method and the weighted voting method can achieve the purpose of this invention, and the specific method to be used can be pre-configured according to the airport security strategy.

[0373] After obtaining the target anomaly level, compare it with the initial processing channel level determined in step S400, and adjust it according to the following rules:

[0374] Determine the processing channel level corresponding to the target anomaly level: the first anomaly level corresponds to the first processing channel (keeping the original level), the second anomaly level corresponds to the second processing channel, the third anomaly level corresponds to the third processing channel, and the fourth anomaly level corresponds to the fourth processing channel.

[0375] If the processing channel level corresponding to the target anomaly level is the same as the initial processing channel level, then the initial processing channel level remains unchanged.

[0376] If the processing channel level corresponding to the target anomaly level is higher than (i.e., the security requirements are more stringent) the initial processing channel level, then the initial processing channel level will be adjusted to the processing channel level corresponding to the target anomaly level.

[0377] If the processing channel level corresponding to the target anomaly level is lower than the initial processing channel level, then the initial processing channel level is maintained unchanged (i.e., the security level is not reduced).

[0378] The core principle of the adjustment rules is that dynamic trajectory calibration is only used to improve the security level (upgrading from a low-intensity channel to a high-intensity channel), and never to reduce the established channel level, ensuring that the overall security intensity is not lower than the baseline level of the initial rating.

[0379] The adjusted processing channel level will be the final channel allocation result for the target user. If the adjusted level differs from the initial level, the user will be redirected to the updated channel; otherwise, the original channel will be used.

[0380] Through the above steps, this invention constructs a two-stage hierarchical classification architecture that first implements broad access and then performs dynamic calibration:

[0381] The first phase (S300-S400) calculates the fusion feature value based on multi-source pre-security inspection data and historical data, establishes the initial channel level, and grants the first processing channel (fast passage) access qualification to users who meet the first preset threshold (e.g., fusion feature value ≥ 85 points), thereby realizing the initial diversion of low-risk users.

[0382] The second phase (S501-S505) targets users who have already obtained fast-pass privileges. It utilizes spatiotemporal trajectory data for segmented clustering comparison. By calculating the high-dimensional feature deviation between the trajectory feature vector and the normal behavior clustering model, a nonlinear mapping from the behavior space to the channel level is established. When the deviation exceeds a preset tolerance threshold, a dynamic upgrade of the channel level is triggered (gradually upgrading from the first processing channel to the second, third, or fourth processing channel), forming a closed-loop control mechanism of "access-monitoring-feedback-adjustment".

[0383] The technical effect of the technical solution provided in this application is as follows:

[0384] 1. Dual-layer heterogeneous feature fusion: Static attribute features (identity identifier, historical record, flight attributes) and dynamic spatiotemporal behavioral features (trajectory geometry, motion attributes, behavioral semantics) are extracted and weighted in layers, overcoming the decision bias problem caused by a single feature source.

[0385] 2. Post-anomaly compensation mechanism: For users whose initial fused feature values ​​meet the first threshold (i.e., low anomaly), secondary verification is performed through high-resolution spatiotemporal trajectory behavior analysis, which makes up for the technical deficiency of the traditional differentiated security inspection mechanism that is based on static features and static binding mechanism of channel level, which cannot respond to real-time behavioral changes.

[0386] 3. Hierarchical Adaptive Decision Making: A continuous quantitative mapping relationship is established between the degree of feature deviation and the channel upgrade level (the greater the deviation, the greater the upgrade). Unlike simple binary (normal / abnormal) judgment, this realizes the fine-grained step-by-step adjustment of security inspection intensity.

[0387] 4. Guarantee of minimum throughput efficiency: Dynamic trajectory calibration is only performed on users initially identified as being in the first processing channel, and the adjustment direction is one-way upgrade (i.e., without downgrading the already determined channel level), ensuring that most users are not affected by additional processing steps and maintaining the throughput efficiency advantage of differentiated security checks.

[0388] Through the above design, this invention establishes a closed-loop verification mechanism based on real-time spatiotemporal trajectory behavior analysis while maintaining the efficiency of fast lane passage, thus achieving technical unification of passage efficiency and security inspection intensity in differentiated security inspection systems.

[0389] Second Embodiment

[0390] This embodiment provides a travel user classification system, including:

[0391] One or more processors;

[0392] Memory, used to store one or more programs;

[0393] When the one or more programs are executed by the one or more processors, the one or more processors implement the travel user classification method as described in the first embodiment.

[0394] Furthermore, the system also includes:

[0395] The data acquisition interface is used to acquire real-time detection data, historical data, and spatiotemporal trajectory data from multiple sources during pre-security checks;

[0396] The communication module is used to interact with the security checkpoint control system and send the determined processing channel level to the channel guidance device.

[0397] This invention also provides a computer-readable storage medium storing computer-executable instructions for performing the methods described in this invention.

[0398] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0399] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for classifying travel users, characterized in that, include: Acquire real-time multi-source detection data, historical data, and spatiotemporal trajectory data of the target user during pre-security checks; Based on the aforementioned pre-security check multi-source real-time detection data, a pre-security check multidimensional dataset of the target user is generated; The pre-security check multidimensional dataset and the historical data are input into a multi-level weighted calculation model to calculate the fusion feature value of the target user. Based on the fusion feature value, the initial processing channel level corresponding to the target user is determined, and the initial processing channel level includes a first processing channel level corresponding to the fusion feature value being greater than or equal to a first threshold. When the initial processing channel level is the first processing channel level, the initial processing channel level is adjusted based on the spatiotemporal trajectory data of the target user, and the adjusted processing channel level is used as the classification result of the target user.

2. The method according to claim 1, characterized in that, The multi-level weighted calculation model includes a first feature layer constructed based on the pre-security check multidimensional dataset, a second feature layer constructed based on user identity and behavior feature data, a third feature layer constructed based on historical risk feature data, a fourth feature layer constructed based on historical travel feature data, and a fifth feature layer constructed based on time and flight attributes. Each feature layer and its sub-features are assigned weight values ​​and quantization rules. When any preset abnormal feature is triggered, the target user is determined to be an abnormal user and subsequent calculations are stopped.

3. The method according to claim 1, characterized in that, The adjustment of the initial processing channel level based on the spatiotemporal trajectory data of the target user specifically includes: The spatiotemporal trajectory data is divided into multiple trajectory segments, and the time information corresponding to each trajectory segment is recorded; Feature extraction is performed on each trajectory segment to generate the corresponding trajectory feature vector; The trajectory feature vector of each trajectory segment is input into a clustering model associated with the time information of that trajectory segment. The output of the clustering model includes at least one first-class cluster result and one second-class cluster result. When the output result is the second type of cluster result, calculate the degree of deviation between the trajectory feature vector of the trajectory segment and the first type of trajectory cluster that is closest to the trajectory feature vector in the feature space, and determine the anomaly level of the trajectory segment based on the degree of deviation; The anomaly level of each trajectory segment is combined to obtain the trajectory anomaly level of the target user, which is used as the comparison result, and the initial processing channel level is adjusted based on the comparison result.

4. The method according to claim 3, characterized in that, Different clustering models correspond to different time information ranges, and the clustering radius threshold of the clustering model corresponding to the trajectory segment with a shorter time interval with the expected event is smaller.

5. The method according to claim 3, characterized in that, The degree of deviation is the ratio of the feature distance between the trajectory feature vector and the nearest first-class trajectory cluster to the cluster radius of the nearest first-class trajectory cluster; The determination of the anomaly level of the trajectory segment based on the degree of deviation specifically includes: When the deviation is less than a first preset threshold, the abnormality level of the trajectory segment is determined to be the first abnormality level, whereby the first abnormality level indicates that the current processing channel level remains unchanged. When the deviation is greater than or equal to the first preset threshold and less than the second preset threshold, the abnormality level of the trajectory segment is determined to be the second abnormality level. The second abnormality level indicates that the processing channel level will be adjusted to the second processing channel level. When the deviation is greater than or equal to the second preset threshold and less than the third preset threshold, the abnormality level of the trajectory segment is determined to be the third abnormality level, and the third abnormality level indicates that the processing channel level is adjusted to the third processing channel level. When the deviation is greater than or equal to the third preset threshold, the abnormality level of the trajectory segment is determined to be the fourth abnormality level, and the fourth abnormality level indicates that the processing channel level is adjusted to the fourth processing channel level. The security inspection intensity corresponding to the first processing channel level, the second processing channel level, the third processing channel level, and the fourth processing channel level increases progressively.

6. The method according to claim 5, characterized in that, At least one of the first preset threshold, the second preset threshold, and the third preset threshold is a dynamic threshold; The dynamic threshold is determined in the following ways: During the pre-security check phase, facial image sequences of the target user are collected, facial expression feature vectors are extracted and input into the facial expression classification model, and the corresponding emotional state category is output. A preset correction coefficient is obtained based on the emotional state category; The dynamic threshold is obtained by multiplying the base threshold by the correction coefficient.

7. The method according to claim 1, characterized in that, The process of determining the initial processing channel level based on the fusion feature value includes: When the fused feature value is greater than or equal to the first feature value and no preset abnormal feature is triggered, the initial processing channel level is determined to be the first processing channel level; When the fusion feature value is less than the first feature value and greater than or equal to the second feature value, the initial processing channel level is determined to be the second processing channel level; When the fusion feature value is less than the second feature value and greater than or equal to the third feature value, the initial processing channel level is determined to be the third processing channel level; When the fusion feature value is less than the third feature value or any preset abnormal feature is triggered, the initial processing channel level is determined to be the fourth processing channel level. Wherein, the first feature value is greater than the second feature value, and the second feature value is greater than the third feature value; the security inspection intensity corresponding to the first processing channel level, the second processing channel level, the third processing channel level, and the fourth processing channel level increases progressively.

8. The method according to claim 5, characterized in that, The step of adjusting the initial processing channel level based on the comparison results specifically includes: The target anomaly level is obtained by comprehensively evaluating the anomaly levels of each trajectory segment. If the processing channel level corresponding to the target anomaly level is different from the initial processing channel level, then the initial processing channel level is adjusted to the processing channel level corresponding to the target anomaly level.

9. The method according to claim 3, characterized in that, The time segmentation rules used to divide the spatiotemporal trajectory data into multiple trajectory segments result in different durations for different trajectory segments, and the longer the time interval between the trajectory segment and the expected event time, the longer the duration of that trajectory segment.

10. A travel user classification system, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the travel user classification method as described in any one of claims 1 to 9.

Images