Control system and control method of anti-backflow channel

Abstract

The invention discloses a control system and a control method of an anti-backflow channel, and relates to the technical field of traffic management of traffic personnel in a safety management and control area. The control method comprises the steps of obtaining a track state and a track association result, judging abnormal behaviors in a channel according to the track state and the track association result, carrying out ID (Identity) sorting, marking and logout on personnel, generating various mode instructions according to the calculated channel passing width, the required channel number and the abnormal behavior judgment result, and various execution terminals in the system are controlled based on various mode instructions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of safe area access management, and particularly relates to a control system and a control method of an anti-backflow passage. BACKGROUND

[0002] The anti-backflow passage system is suitable for security level classification and control areas such as airport restricted areas, customs passages and data centers, and combines physical interception and intelligent algorithms to prevent personnel from entering high-security areas from low-security areas, and supports unattended operation. The system is composed of a sensing layer, a decision layer, an execution layer and a monitoring layer, can monitor the movement direction of personnel, identify retained objects, detect the trajectory of thrown objects and the like, and upload information to a management platform.

[0003] Currently, the system has many technical defects: low accuracy of abnormal behavior determination, traditional systems rely on a single sensor, have insufficient ability to determine complex abnormal behaviors and have a high misjudgment rate; there is a prominent contradiction between energy consumption and response speed, the door control system is often powered on and driven, has high energy consumption and slow response; poor flow adaptability, fixed door control mode is adopted and cannot be dynamically adjusted; lack of multi-abnormality cooperative detection capability, most solutions only prevent reverse movement and have insufficient detection of other security risks; real-time performance is insufficient due to reliance on cloud processing, and network delay affects abnormality interception capability.

[0004] In view of these pain points, an anti-backflow technical solution needs to be fused with multi-modal sensing hardware, layered decision architecture and efficient data processing algorithms to improve recognition accuracy, balance energy consumption, response speed and flow adaptability, and meet the needs of full-scene application. SUMMARY

[0005] According to a first aspect of some embodiments of the present application, a control method of a reflux-preventing passage system is provided, comprising: obtaining an updated personnel trajectory state, an object trajectory state and a trajectory association result according to obtained personnel depth map centroid coordinates, object depth map centroid coordinates, and a penetration point coordinate and a historical trajectory state, the trajectory association result comprising at least one of a successful association, a new trajectory generation, and a trajectory termination; determining a reverse behavior, a loitering behavior, a throwing behavior and a left-behind object in the passage based on the personnel trajectory state, the object trajectory state and the trajectory association result, ReID features, obtained light curtain timing data, and obtained camera pose data, to obtain an abnormal behavior determination result; assigning, marking and canceling the ID of a passing person through an improved ReID algorithm according to a depth feature data source, an appearance feature data source collected by a dual-spectrum camera, the personnel trajectory state and the abnormal behavior determination result; calculating a passage passing width and a required number of passages; and generating at least one of a flow mode instruction, a free mode instruction, an interlocking mode instruction, a locking mode instruction, a maintenance mode switching instruction, an emergency mode switching instruction, a high-risk abnormality corresponding instruction, a medium-risk abnormality corresponding instruction, and a low-risk abnormality corresponding instruction based on at least one of the passage passing width, the required number of passages, and the abnormal behavior determination result.

[0006] According to a second aspect of some embodiments of the present application, a control system of a reflux-preventing passage is provided, comprising a trajectory state tracking module, an abnormal behavior determination module, an ID module, a passage calculation module, and an instruction control module. The control system is configured to perform the control method of the reflux-preventing passage system according to the first aspect.

[0007] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and are not limiting to the present application. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 A flowchart of a control method of a reflux-preventing passage in some embodiments is shown.

[0009] Figure 2 A flowchart of step S10 in some embodiments is shown.

[0010] Figure 3 A flowchart of step S20 in some embodiments is shown.

[0011] Figure 4 A flowchart of step S30 in some embodiments is shown. DETAILED DESCRIPTION

[0012] The present application relates to a control system and a control method of a backflow prevention channel, which is suitable for security level grading management and control scenes such as airport restricted areas, customs channels, data centers and the like, and aims to realize the prevention and control of reverse intrusion from a low security area to a high security area through physical interception and intelligent algorithm cooperation, while supporting unattended operation and adaptive flow operation.

[0013] In the entity hardware application scene, the present application provides a backflow prevention channel control system, which comprises a door control, a detection and sensing, a control decision and a communication transmission element. The door body assembly comprises a high security A door and a low security B door, which are constructed by 30mm thick acrylic / PC explosion-proof materials. The detection and sensing assembly comprises a TOF sensor array, a dot matrix light curtain, a dual-spectrum camera and a pressure sensor. The TOF sensor array has a wavelength of 940nm, a frame rate of greater than or equal to 60fps, and an error of less than or equal to 2%. The dot matrix light curtain is partially erected on both sides of the channel and partially embedded in the ground. The dual-spectrum camera integrates two kinds of pixel sensors, has automatic white balance and noise reduction functions, has a resolution of 4K@30fps, and has an infrared detection distance of 0.5-10m. The pressure sensor is embedded in the edge of the door body, has a trigger force of less than 50N, and has a response time of less than or equal to 10ms. The control decision element integrates a processor and a GPU, has built-in memory and storage, has a computing power of 200TOPS, has a local processing delay of less than 50ms, and supports offline algorithms. The door control execution element is composed of an electromagnetic lock body and the like, integrates a position and an overcurrent protection module, has a response time of less than 0.3s, and is normally powered off and locked. The communication transmission element supports 5G and Ethernet, integrates an interface, and supports VPN encryption. The detection and sensing assembly is connected to the control decision element through an industrial Ethernet, and has a transmission delay of less than or equal to 20ms. The control decision element is connected to the door control execution element through an RS485 bus, issues instructions and triggers an alarm, is connected to the communication transmission element through a PCIe interface, uploads channel state information and receives platform instructions, and receives a feedback anti-pinch signal from the electromagnetic door pressure sensor, interrupts the instructions in the control decision element, and reverses the door.

[0014] The present application also provides a control method of a backflow prevention channel, which collects data of passing personnel and / or articles in the channel through dynamic sensing, tracks the motion trajectory of the passing personnel and / or articles, determines the abnormal behavior of the passing personnel, and marks the personnel and calculates the space parameters to form the mode decision and response result of the backflow prevention channel control.

[0015] Figure 1 A flowchart of the control method of the backflow prevention channel in some embodiments is shown. As shown in Figure 1 The control method of the backflow prevention channel can comprise the following steps S10-S50.

[0016] In step S10, according to the obtained personnel depth map centroid coordinates, object depth map centroid coordinates, and penetration point coordinates, and the historical trajectory state, the updated personnel trajectory state, object trajectory state, and trajectory association result are obtained, the trajectory association result including at least one of association success, new trajectory generation, and trajectory termination. In this step, by fusing the detection data of the TOF sensor and the dot matrix light curtain, a motion state model of the personnel / object is established, continuous trajectory tracking of multiple targets is realized, the problems of trajectory loss and confusion in the occlusion and intersection scenarios are solved, and accurate motion data support is provided for subsequent abnormal behavior judgment.

[0017] Figure 2 A flowchart of step S10 in some embodiments is shown. Specifically, step S10 includes:

[0018] Step S11: According to the collected personnel depth map centroid coordinates, object depth map centroid coordinates, and penetration point coordinates, data fusion is performed to obtain the personnel fusion observation value vector and the observation value vector.

[0019] Specifically, step S11 includes:

[0020] Step S111: Obtain the personnel depth map centroid coordinates, object depth map centroid coordinates, and penetration point coordinates synchronized in sampling frequency; Specifically, the personnel / object depth map centroid coordinates (X tof ,Y tof ) collected by the TOF sensor, the penetration point coordinates (X light ,Y light ) collected by the lateral light curtain, and the sampling frequency are synchronized with the frame rate of the TOF sensor (60HZ);

[0021] Step S112: Determine the personnel motion state vector or object motion state vector according to the personnel depth map centroid coordinates, object depth map centroid coordinates, and penetration point coordinates, the motion state vector being X x ,v y ,a x ,a y ] T , wherein x and y are two-dimensional positions (unit: m), v x ,v y are horizontal and vertical velocities (unit: m / s) respectively, and a x ,a y are horizontal and vertical accelerations (unit: m / s²) respectively;

[0022] Step S113: Establish a discrete-time state equation , wherein X t is the motion state vector of t frame, X t-1 is the motion state vector of t-1 frame, F is a state transition matrix, From The process noise of motion disturbance during the time interval from frame 1 to frame t; the state transition matrix F = [10Δt00.5Δt] 2 0;010Δt00.5Δt 2 ;0010Δt0;00010Δt;000010;000001];Δt=1 / 60s is the sampling interval, That is, the process noise vector follows a multivariate Gaussian distribution with a mean of 0 to a covariance matrix of Q; Q is the process noise covariance matrix, with diagonal elements of 0.01, 0.01, 0.005, 0.005, 0.001, and 0.001 respectively.

[0023] Step S114: Establish the observation equation as Z t =HX t +V t Among them, Z t Let X be the observation vector of frame t. t Let V be the motion state vector of frame t, H be the observation matrix, H=[100000;010000], used to observe position information, and V be the motion state vector of frame t. t The observation noise vector follows a multivariate Gaussian distribution with a mean of 0 to a covariance matrix of R; R is the observation noise covariance matrix, and the diagonal elements of R are set to 0.0025 and 0.0025 according to the TOF and light curtain precision.

[0024] Step S115: Obtain the personnel observation vector or object observation vector based on the discrete-time state equation and observation equation. The observation vector is Z. t =[Z t , x Z t,y ] T Among them, Z t,x =ω tof x tof +ω light x lightt Z t,y =ω tof y tof +ω light y light Z t , x Z t,y To determine the two-dimensional position coordinates of a person or object in the channel plane based on the observation results, ω tof ω represents the centroid coordinate weight of the depth map. light For the weight of the penetration point coordinates; in some embodiments, ω tof Setting it to 0.7 indicates that TOF has higher positional accuracy and a larger weight, ω light A value of 0.3 indicates that the light curtain is used for supplementary verification.

[0025] This step establishes a mathematical model that conforms to the movement law of personnel / objects based on the centroid coordinates and penetration point coordinates of the collected depth map of personnel / objects, reducing noise interference from a single sensor and reducing the position observation error from ±10cm to ±5cm.

[0026] Step S12: Based on the fused observations of the current frame, historical trajectory status, and ReID feature similarity, obtain trajectory association results. The trajectory association results include at least one of the following: successful association results, generated new trajectory, updated trajectory, and trajectory termination results.

[0027] Specifically, step S12 includes:

[0028] Step S121: Obtain the historical trajectory state using the Kalman prediction method. The historical trajectory state is estimated by filtering the values ​​of frame t-1. and determination of covariance matrix Prior covariance . The mid-coordinate represents the predicted position of the historical trajectory in the current frame. Quantify the uncertainty of the predicted location. for The output of frame execution step S12 is the system's "optimal estimate" of the historical motion state, forming an iterative connection in time sequence. If it is the first frame of data after the system starts, there is no historical trajectory to inherit, and it needs to be generated through "initial state setting": Filtered estimate x^0|0: Using the first frame fused observation vector Z0 output by sub-step S11 as the position reference, the velocity and acceleration are initialized to 0, i.e., x^0|0=[Z0,x,Z0,y,0,0,0,0] T Covariance matrix P0|0: Initialized as the observation noise covariance matrix R (diagonal elements [0.0025, 0.0025, 0.01, 0.01, 0.005, 0.005]), representing the uncertainty of the initial state (position error is determined by sensor accuracy, velocity / acceleration error is set to empirical values); after each frame is completed ( The "Data Association and Trajectory Update" for each frame will output the optimal filtered estimate for that frame. Covariance Matrix The data is stored in the cache of the edge computing unit and used as historical trajectory data for the next frame (frame t). The historical trajectory state serves as the "historical baseline" for trajectory prediction in the current frame, and the prior state of the current frame is obtained recursively through the state transition matrix F. It is the core input of the "prediction-update" closed loop.

[0029] Step S122: Extract a person recognition feature from the collected image, obtain a ReID feature vector, the ReID feature vector includes a current frame feature and a historical trajectory feature, and determine a ReID feature similarity according to the current frame feature and the historical trajectory feature; the ReID feature similarity S ReID,ij t t-1 t t-1 t t-1 is respectively a current frame feature and a historical trajectory feature, the ReID feature similarity is 0~1, and the closer to 1, the higher the matching degree is;

[0030] Step S123: Calculate a correlation cost of a current observation target and a historical trajectory;

[0031] For each current observation target i (corresponding to a fusion observation value Z t,i ) and each historical trajectory j (corresponding to a prior prediction position X^ t∣t-1,j ), a correlation cost C ij is calculated, wherein α is a position distance weight (for example, 0.7 is taken), β is a ReID feature weight (for example, 0.3 is taken), i is a current observation target index, and j is a historical trajectory index;

[0032] Step S124: Determine whether to match the observation target of the current frame and the historical trajectory according to the correlation cost;

[0033] Step S125: If it is detected that the correlation cost of a certain observation target is less than a cost matrix threshold, and the correlation cost is the minimum cost of the observation target and the minimum cost of the historical trajectory, it is determined that the correlation is successful; in some embodiments, the cost matrix threshold is 0.5;

[0034] Step S126: If it is detected that the correlation costs of multiple historical trajectories and the same observation target are less than the cost matrix threshold, the historical trajectory corresponding to the minimum correlation cost is selected as the correlation result;

[0035] Step S127: If it is detected that all correlation costs of a certain observation target are not less than the cost matrix threshold, and the number of current frame observation targets is greater than the number of historical trajectories, it is determined that a new trajectory is detected, and the trajectory state (x^ t∣t = Z t,i , P t∣t =R) of the target is initialized and a new trajectory ID is assigned;

[0036] Step S128: If it is detected that the prior covariance ​​​​​​If the covariance exceeds the preset threshold, a step of judging the consistency of the moving direction of the current observation target and the velocity vector direction of the historical trajectory is performed; if the prediction uncertainty is high due to occlusion, additional verification (the included angle is less than 30°) is required to avoid false association;

[0037] Step S129: If the association is successfully detected, the posterior covariance P is obtained through Kalman update t∣t If the covariance still exceeds the preset threshold (more than twice the initial covariance), the trajectory is marked as “to-be-confirmed association”, and the next frame needs to be verified; Kalman update: after successful association, the Kalman gain K is calculated , the updated state , and the updated covariance .

[0038] Further, the step S12 further includes a judgment step of trajectory termination. The trajectory termination refers to determining that the target corresponding to a historical trajectory has left the channel or disappeared (such as a person walking out of the detection range or being unable to recover after being blocked), and the historical data of the trajectory needs to be deleted to avoid redundancy.

[0039] Specifically, the step S12 further includes: recording the association state of the trajectory; judging the time sequence interruption and the occlusion timeout; determining whether to trigger the trajectory termination according to the coordinates of the spatial position; and performing a trajectory termination operation. The trajectory termination operation includes: deleting the historical data of the trajectory (including , the corresponding features in the ReID feature library; if the trajectory is associated with an abnormal behavior (such as a reverse behavior or a parabolic behavior), the ID and the abnormal record are retained for subsequent tracing; updating the trajectory list, moving the terminated trajectory from the “active trajectory” list to the “historical trajectory” archiving list, and automatically deleting after 24 hours.

[0040] For each historical trajectory, the “frame number of the last successful association” t last is recorded in real time: if the association is successful in a certain frame, the current frame number is updated; if there is no associated observation target in a certain frame, it is marked as “interruption frame”, t last remains unchanged; if the feature is invalid due to occlusion, it is marked as “to-be-confirmed association”, t last remains unchanged, but the “to-be-confirmed starting frame number” t pending is additionally recorded; the difference between the current frame number t and t last is calculated: if Δt≥3 frames (preset termination frame threshold), and the association cost of any observation target is less than the cost matrix threshold, it is determined that the time sequence is interrupted and terminated; if the trajectory is in the “to-be-confirmed association” state, Δtpending is calculated: : if Δtpending≥10 frames, and there is no depth signal of the TOF sensor at the predicted position of the trajectory, it is determined that the occlusion timeout is terminated.

[0041] Prior predicted position based on historical trajectory In the x coordinate (channel length direction) in , if the coordinate of the person in the channel length direction exceeds the channel entrance end or exceeds the channel exit end, and no observation target is matched for 2 frames, and the bottom light curtain and the dual-spectrum camera do not detect the target, the trajectory termination is triggered; if the predicted position of the target in the coordinate of the channel length direction is between the channel entrance end and the exit end, but the depth signal and the light penetration point signal cannot be obtained for two consecutive frames, it is determined that the "abnormal disappearance" is triggered, and the trajectory termination is triggered; so as to avoid the redundant trajectory caused by temporary failure of the sensor.

[0042] After this step, the motion trajectories of multiple observation targets in the cross scene are associated and matched, the accuracy is improved from 85% to 98%, and the success rate of trajectory reconnection after occlusion recovery is > 95%.

[0043] In step S20, based on the personnel trajectory state, the object trajectory state, the trajectory association result, the ReID feature, the obtained light curtain timing data and the obtained camera posture data, the reverse behavior, the wandering behavior, the throwing behavior and the left object in the channel are determined, and the abnormal behavior determination result is obtained; four types of abnormal behaviors are calibrated and recognized based on the historical data of the airport and the customs, so as to ensure consistency.

[0044] Figure 3 A flowchart of step S20 in some embodiments is shown. In some embodiments, step S20 includes determining the reverse behavior of the person in the channel based on the consistency result of the direction of the person's torso and the direction of movement, and the direction of the channel safety level (S21), which can accurately identify the reverse intrusion behavior and exclude normal side turning and other misjudgment scenes.

[0045] Specifically, step S21 includes:

[0046] Step S211 includes normalizing the velocity vector direction in the trajectory state vector of the person, so as to match the torso orientation angle dimension;

[0047] The speed vector direction θ ranges from 0° to 360° (including reverse 180° to 360°), and the trunk orientation angle ϕ ranges from 0° to 180°. θ needs to be normalized to 0° to 180° to ensure that they can be directly compared. When θ ∈ [0°, 180°], θ' = θ, the direction is consistent with ϕ, and no adjustment is needed; when θ ∈ (180°, 360°], θ' = 360°-θ, the reverse speed direction is mapped to the positive angle, such as θ = 270°→θ' = 90°, indicating that the speed direction and the trunk orientation are transversely angled at 90°; if θ is "invalid" (speed < 0.1 m / s), skip the normalization and directly mark it as "to be confirmed", and wait for the next frame of valid speed data to process. Through normalization, the angle calculation between "speed direction and trunk orientation" is from "0° to 180°" single interval comparison, which avoids the angle calculation error caused by the 360° cyclic characteristics of θ, such as θ = 350° and ϕ = 10°, the angle is 340° without normalization, and it is 20° after normalization, which conforms to the actual motion logic.

[0048] Step S212 includes: if the trunk orientation angle is less than 30°, mark it as the first candidate reverse behavior; if the absolute value of the angle difference between the speed vector direction and the trunk orientation angle is less than 15°, mark it as the second candidate reverse behavior; if the trunk orientation angle is less than 30° and the absolute value of the angle difference between the speed vector direction and the trunk orientation angle is greater than 15°, mark it as an abnormal posture behavior;

[0049] Among them, the first candidate reverse behavior indicates that the trunk orientation of the person in the channel is in the low safety zone, which is opposite to the positive direction of the channel, indicating that the subjective intention of the person may be reverse. The second candidate reverse behavior indicates that the movement direction of the person in the channel is consistent with the trunk orientation, indicating that the person is "actively moving towards the low safety zone", rather than passively adjusting the posture. The abnormal posture behavior indicates that the trunk orientation of the person is in the low safety zone, but the speed direction is towards the high safety zone, which is determined as "abnormal posture" and not marked as "reverse candidate". When marked as "reverse candidate", the edge computing unit triggers the lateral light curtain "encryption scanning" of the direction recognition layer (from 5 lines / cm to 8 lines / cm), which verifies the movement direction of the person through the transverse penetration track of the light curtain to avoid the calculation error of θ caused by the TOF sensor due to shielding; at the same time, the dual-spectrum camera automatically enlarges the trunk region of the person (ROI is scaled to 1.5 times of the original size), which improves the calculation accuracy of the next frame of ϕ.

[0050] Step S213 includes maintaining some reverse candidate counters for each trajectory, when a certain frame meets the "reverse candidate" condition, the count is incremented by 1; when a certain frame does not meet, the counter is cleared (to avoid cumulative errors); when the counter is greater than or equal to 3, the displacement verification step is entered; if a frame or invalid (such as occlusion), the counter is suspended, and after the data is recovered, it continues, avoiding verification interruption caused by temporary data loss. "Single frame reverse candidate" may be caused by instantaneous action (such as personnel suddenly turning around but not moving, limb swinging to block the sensor), which needs to be excluded by "multi-frame continuous verification".

[0051] Step S214 includes calculating the reverse displacement based on the displacement coordinates in the channel length direction in the trajectory association result, and if the reverse displacement is greater than the displacement threshold, it is marked as a third candidate reverse behavior. Wherein, the reverse displacement is the displacement in the previous 3 frames is the position 3 frames ago, x t is the current frame position, the displacement threshold is 3 cm, and the third candidate reverse behavior indicates that the personnel moves at least 3 cm in 3 frames (0.05s) to the low safety area, the speed is greater than or equal to 6m / s, and there is a reverse tendency.

[0052] Step S215 includes screening the verified reverse behavior from the first candidate reverse behavior, the second candidate reverse behavior, and the third candidate reverse behavior; specifically, screening according to the reverse confidence, the reverse confidence is , and Δθ is the absolute value of the angle difference between the velocity vector direction and the trunk orientation angle. The smaller Δθ is, the higher the confidence is. For example, when Δθ = 0°, the confidence = 1.0; when Δθ = 15°, the confidence = 0.85. When the confidence is greater than 0.9, it is determined as "high confidence reverse" (such as personnel quickly and clearly moving to the low safety area); when the confidence is 0.85-0.9, it is determined as "medium confidence reverse" (such as personnel slowly and hesitantly moving to the low safety area), which provides a hierarchical basis for subsequent abnormal response. The reverse misjudgment rate is reduced from 5% of the traditional scheme to 0.3%, avoiding misinterception of normal queuing personnel.

[0053] In some personnel normal side avoidance scenarios, the channel is bidirectional, the forward personnel temporarily side to avoid the reverse personnel, the trunk is oriented to the low safety area (φ = 25°), but the speed direction is still to the high safety area (θ' = 170°); , which does not meet the "direction consistency" condition and is not marked as "reverse candidate", which excludes misjudgment.

[0054] In some scenarios of temporary personnel turning to take things, the forward moving personnel (ϕ = 100°) suddenly turns to take the back item, the trunk direction becomes ϕ = 20°, but the speed direction is 0 (static, θ invalid). θ invalid, skip normalization, "reverse candidate counter" does not accumulate, and continues to move forward after the personnel turn (ϕ returns to 100°, θ valid), not determined as reverse, to exclude false positives.

[0055] In some scenarios of personnel slowly moving backward (such as looking for missing items), the personnel have entered the high security area (displacement coordinate symbol in the direction of the passage length), found that the item is missing in the low security area, slowly moved to the low security area ( ), the trunk direction is to the low security area ( ). Δθ = |θ | = 20° (close to 15° threshold), but 3 consecutive frames satisfy and Δθ < 20° (the direction consistency threshold in this scenario can be temporarily adjusted to 20° through the remote platform), and 3 frames of displacement D 逆行 = 1.5 cm (≥ 3 cm correction threshold), determined as "medium confidence reverse", triggered A door delay 2s lock, give the personnel time to take back after returning, balance security and humanization.

[0056] As shown in Figure 3 , step S20 further includes quantifying the uncertainty of the motion direction by calculating the information entropy of the direction angle of the personnel trajectory, and identifying the wandering behavior without clear intention of moving (step S22).

[0057] Specifically, step S22 includes:

[0058] Step S221: Discretize the velocity vector direction to obtain a plurality of direction angle intervals;

[0059] The velocity vector direction θ is a continuous value (0°~180°), which cannot be directly calculated as a probability, and needs to be discretized into a limited interval by binning, while considering "resolution" and "statistical effectiveness". Too fine interval will result in no sample in some intervals (probability is 0, entropy value calculation bias), and too coarse interval will lose the direction fluctuation details.

[0060] Divide 0°~180° into 9 intervals, each interval width is 20°, traverse 10 frames of direction angle, and count the number of angles n k(If θ1 falls into interval 2, n2 increases by 1; if θ2 falls into interval 8, n8 increases by 1. If there is no valid angle in a frame (e.g. a static fill value), it is not counted to avoid interference with the statistics. If the number of valid angles in 10 frames is less than 8 (e.g. severe occlusion), the current determination is suspended, and the next sequence of 10 frames is waited for to ensure sufficient sample size. By dividing into 9 intervals, the details of direction fluctuation are preserved (e.g. an interval width of 20° can distinguish between "small adjustment" and "large turn"), and sufficient statistics are ensured for each interval in normal / wandering samples. Normal samples are more concentrated in the middle intervals 5-7, and wandering samples are scattered in each interval, providing a guarantee for the accuracy of subsequent entropy calculation.

[0061] Step S222: Calculate the direction angle probability of each direction angle interval, which is calculated according to the formula , nk is the number of angles in the kth interval, is the total number of valid angles in 10 frames, ≤10, and the sum of the direction angle probabilities of all direction angle intervals is equal to 1;

[0062] wherein, if the total number of valid angles in 10 frames is 0, i.e. all frame angles are invalid, it is determined as "to be confirmed", and the probability is not calculated to avoid mathematical errors with a denominator of 0. If n k =0 (no angle falls into it), p(θk)=0, but special processing is required in the subsequent entropy calculation (0xlog20 is counted as 0, which conforms to the definition of information theory). If all angles are concentrated in one interval, i.e. n k =10, p(θk)=1, and other p(θi)=0, the entropy H=0, which corresponds to "direction completely stable" (normal passage).

[0063] Step S223: Calculate the trajectory entropy according to the direction angle probability, and determine the wandering behavior according to the trajectory entropy, the dwell time and the moving distance;

[0064] The trajectory entropy H is used to measure the "uncertainty of the sequence of direction angles": the smaller H is, the more concentrated the direction is (e.g. when normal passage, the angles are mostly concentrated in 1-2 intervals), and the uncertainty is low; the larger H is, the more dispersed the direction is (e.g. when wandering, the angles fluctuate in multiple intervals), and the uncertainty is high.

[0065] Specifically, ; when all p(θk)=0 or only one p(θk)=0, H=0; when p(θk) is uniformly distributed (e.g. 9 intervals each accounting for 1 / 9), H=log29≈3.17, which is the theoretical maximum value in the passage scenario. During the calculation, if p(θk)=0, then = 0, avoid the logarithm unintentional; entropy value calculation result is kept 2 decimal places (such as H = 1.85), which meets the requirement of judgment accuracy, and reduces the storage and calculation amount; if the entropy value of a frame suddenly changes (such as from 1.0 to 2.5), the moving average of the entropy values of the previous 3 frames (H 平滑 = 0.5H 当前 + 0.3H 前1 + 0.2H 前2 ) is used to avoid misjudgment caused by instantaneous noise.

[0066] Step S224: If the trajectory entropy of the person in the channel is greater than 1.8, the stay time is greater than 5s, and the moving distance is less than 1m, the behavior of the person in the channel is judged as loitering behavior, and the trajectory entropy, stay time and moving distance are recorded;

[0067] Step S225: If the trajectory entropy of the person in the channel is greater than 1.8, the stay time is greater than 5s, and the moving distance is equal to 1.2m, the behavior of the person in the channel is marked as "to be re-determined", and the marking of "to be re-determined" is used to trigger the step of determining the loitering behavior according to the trajectory entropy, stay time and moving distance in the next frame.

[0068] Step S226: If the trajectory entropy of the person in the channel is not greater than 1.8, the stay time is not greater than 5s, and the moving distance is not less than 1m, or one of the trajectory entropy of the person in the channel is not greater than 1.8, the stay time is not greater than 5s, and the moving distance is not less than 1m, the behavior of the person in the channel is determined as normal behavior, and the loitering determination counter is reset.

[0069] As shown in Figure 3 , step S20 further includes determining the throwing behavior of the person in the channel according to the captured motion characteristics of the object and the action characteristics of the person (step S23). In the anti-backflow channel system, accurate determination of the throwing behavior is crucial to ensure the safety of the channel and maintain orderly passage. This determination link comprehensively utilizes the timing data of the light curtain sensor and the human body posture recognition result to construct a multi-dimensional determination system to accurately identify the abnormal behavior of throwing.

[0070] Specifically, step S23 includes:

[0071] Step S231: According to real-time analysis of the timing data output by the light curtain sensor, it is detected whether there is an abnormal light shielding event, and according to the timestamp of the light shielding event and the sensor position information, the motion trajectory of the object is calculated by using least square method and other algorithms;

[0072] The light curtain sensors are arranged along the two sides of the passage at equal intervals (every 20 cm) to form a light curtain array. When an object passes through the light curtain, it will block the infrared light, triggering a change in the sensor output signal. The light curtain sensor continuously scans at a frequency of 100 Hz, recording the timestamp and position information of the light blockage. When multiple light curtain sensors are blocked in a specific order within a short period of time, and the blocking position shows rapid displacement changes, it is preliminarily determined as a possible object crossing event.

[0073] Step S232: Extract the posture features closely related to the parabolic behavior from the skeletal point data, and use the sliding window technique to analyze the posture data of multiple consecutive frames (such as 5-10 frames) to determine the trend of posture changes.

[0074] The dual-spectrum camera (visible light + infrared, 4K@30fps) continuously captures images of personnel in the passage. The OpenPose algorithm based on deep learning is used to detect skeletal points in the image, identifying 17 key skeletal points such as the head, shoulders, elbows, and hands. Through the relative position relationship of these skeletal points, the posture information of the human body is calculated, such as the stretching angle of the arms and the twisting angle of the body. The attention mechanism is introduced to make the algorithm more focused on the upper body area of the personnel, improving the accuracy of skeletal point detection. At the same time, combined with infrared image information, personnel posture can also be stably identified in low light environments, ensuring that high-quality posture data can be provided for parabolic behavior determination in various scenarios.

[0075] The angle between the arm and the body, the bending angle of the elbow, the rotation angle of the wrist, etc. are calculated. At the same time, the relative position change of the head and the body is analyzed to determine whether the personnel have the behavior of looking at the parabolic direction. If the arm is quickly stretched from the contracted state within a short period of time, and the body has a significant twisting action, which conforms to the posture change pattern of the parabolic behavior, it is marked as suspicious posture.

[0076] Step S233: Time align the light curtain time series data and human body posture data to ensure that the two kinds of data at the same time point can be accurately associated, and use the weighted fusion method to fuse the features obtained by light curtain time series analysis and the features obtained by human body posture analysis.

[0077] Specifically, by establishing a time index, the abnormal blocking events detected by the light curtain are matched with the corresponding human body posture data, and the correlation between the two is analyzed.

[0078] Step S234: If the light curtain detects an object crossing event, and the object's movement speed exceeds 2 m / s, or the time interval of the object crossing the light curtain is less than 0.3 s, mark the event as a suspicious parabolic event.

[0079] Step S235: If the human posture recognition result shows that the arm stretching angle exceeds 50 degrees, and the angle changes more than 30 degrees within 0.5 seconds, and the twist angle exceeds 20 degrees, mark the posture as a suspicious throwing posture;

[0080] Step S236: Check whether there is a suspicious throwing posture that matches the suspicious throwing event at the corresponding time point. If the light curtain detects an object crossing event at the same time, the person's posture is in the throwing preparation or execution state, and both have consistency in time and space (such as the object movement direction and the person's arm stretching direction are consistent), further increase the decision weight of the throwing behavior;

[0081] Step S237: Compare the throwing behavior confidence calculated by multi-source data fusion with the preset confidence threshold (such as 0.7). If the confidence is greater than the threshold, it is determined as a throwing behavior; if the confidence is less than the threshold, but within a certain range (such as 0.5-0.7), it is marked as a pending throwing behavior, which needs to be judged again combined with the data of subsequent frames.

[0082] Further, when there are many people in the channel, mutual occlusion occurs, causing inaccurate light curtain or posture recognition data. A method based on a probability model is used to recover and correct the data.

[0083] Further, for some complex behaviors similar to throwing behavior, such as waving hands to greet, arranging clothes, etc., distinguish them by analyzing the duration, motion amplitude and frequency of the behavior. Throwing behavior usually has the characteristics of fast, short and large motion amplitude, while waving hands to greet and other behaviors are relatively slow, small in amplitude and have certain regularity. By setting corresponding time, amplitude and frequency thresholds, the throwing behavior can be accurately identified to avoid confusion with other normal behaviors.

[0084] In some reflux channel application scenarios, the time, place, and personnel identity information (if identified) of the parabola behavior, light curtain timing data, human posture data, and determination results are stored to form detailed behavior records. These data are stored in the local edge computing unit and uploaded to the remote management platform regularly for subsequent query and analysis. In conjunction with the monitoring cameras in the channel, when a parabola behavior is detected, the video clips of a certain period of time (such as 10s before and 20s after) before and after the event are automatically marked and saved. Security personnel can view these videos at any time through the remote management platform to more intuitively analyze and trace the parabola behavior, providing strong evidence for subsequent security management and event handling. The stored parabola behavior data is regularly analyzed to summarize the occurrence regularity, common scenarios, and personnel characteristics of parabola behavior. Through data analysis, it is found that the weak links in channel security management, such as a higher incidence of parabola behavior in some areas, may be due to unclear signs or insufficient personnel guidance in that area. According to the analysis results of abnormal behavior, the security management strategy is adjusted accordingly. At the same time, the characteristics and determination method of parabola behavior are included in the employee training content to improve the identification and response capabilities of security personnel to this type of abnormal behavior, preventing the occurrence of parabola behavior from the source and improving the overall safety of the channel.

[0085] As shown in Figure 3 Step S20 further includes identifying the abandoned object according to the comparison result (step S24) according to the continuous occlusion detection result and the static profile. This step accurately identifies the luggage, packages, and documents (minimum identification volume > 5 cm³) left in the channel by fusing the light curtain interruption data and TOF (Time of Flight) profile analysis results, while avoiding misjudging ground stains and temporary trampled objects as abandoned objects. This step realizes efficient and low-misjudgment identification of abandoned objects through the double-verification logic of "bottom light curtain interruption detection (judging the length of stay) + TOF depth profile analysis (confirming the existence of objects)", combined with the specific parameter thresholds and algorithm models in the technical solution.

[0086] Specific steps S24 include:

[0087] Step S241: generate a preliminary abandoned object screening result containing a light curtain candidate region identification result through light curtain detection data; to exclude short-term interference and generate an abandoned object candidate region;

[0088] Specifically, step S241 comprises: monitoring the light curtain data stream in real time, marking a detection point as a “suspicious interruption point” when the interruption time of the detection point exceeds the interruption time threshold and the corresponding object height exceeds 10 cm; defining a region composed of the adjacent three or more detection points as a “light curtain candidate region” and recording the ground coordinates and area of the region (for example, the area of a region composed of 3*3 detection points = 9*(2cm*2cm) = 36cm2) if the adjacent three or more detection points meet the triggering condition of the “suspicious interruption point”; determining that the corresponding object is a “ground litter” if there is only a single “suspicious interruption point” and the corresponding object height is not higher than 15 cm, and not generating a candidate region; determining that it is “temporary stay of personnel” if the interruption time is between 5-8s and the moving distance of the personnel in the channel is greater than 10 cm, and resetting the interruption timing; outputting the coordinates, area and interruption time of the light curtain candidate region, and terminating the process if there is no suspicious region.

[0089] Step S242: confirming whether the light curtain candidate region has a real object protrusion through TOF depth map comparison, generating a contour verification result containing a protruding pixel region identification result; avoiding misjudgment of the light curtain caused by ground stains and shadows (the technical solution can reduce the misjudgment rate by 90% in this link).

[0090] Specifically, step S242 comprises: removing the background depth map from the obtained TOF depth map to obtain a depth difference map; denoising the depth difference map; extracting pixel points exceeding a depth difference threshold from the denoised depth difference map to form a protruding pixel region, calculating the ground projection area of the protruding pixel region; and calculating the center coordinates of the protruding pixel region and generating a depth contour of the protruding pixel region if it is detected that the ground projection area of the protruding pixel region exceeds a protrusion area threshold. The denoising process can use 3*3 Gaussian filtering for smoothing processing to remove the random noise of the TOF sensor (such as pixel points with a depth value jump of <3cm) and retain the depth difference features of real objects. The depth difference is , D i (x,y) is the pixel coordinate vector in the TOF depth map, D bg (x,y) is the pixel coordinate vector in the background depth map, the depth difference threshold can be 5cm, and the protrusion area threshold can be 5cm 2 .

[0091] Step S243: determining whether the light curtain candidate region and the protruding pixel region are left objects through spatial correlation.

[0092] Specifically, step S243 includes: calculating the Euclidean distance between the center coordinates of the light curtain candidate region and the center coordinates of the raised pixel region; if the Euclidean distance is less than the channel sensor deployment accuracy, calculating the overlapping area ratio of the light curtain candidate region and the raised pixel region; if it is detected that the overlapping area ratio exceeds 80%, determining that the association is successful, and determining that the light curtain candidate region and the raised pixel region are the same object; if it is determined that the association is successful, and it is detected that the raised pixel region is greater than the raised area threshold and the interruption time exceeds the interruption time threshold, it is determined to be a residual object; if it is not determined that the association is successful, the light curtain detection data and the TOF depth map after 3 seconds are obtained, and the residual object determination step is performed again; outputting the residual object determination result, if the determination result contains a residual object, the position information and the volume of the residual object are also outputted, the position information includes the light curtain region coordinates, the raised pixel region coordinates, and the coordinate average of the light curtain region and the raised pixel region.

[0093] In some scenarios where personnel gathering causes the light curtain to be triggered, when the TOF sensor detects that there is personnel in the "suspicious residual object area" (the depth value is between 1.5-2.2m, corresponding to the height range of personnel), even if the light curtain meets the interruption condition, the determination is suspended, and the timing is restarted after the personnel leave (TOF depth value returns to stable). That is, when personnel are present, the TOF depth value fluctuation range is 1.5-2.2m, which is obviously distinguished from the object depth value (0.3-1.0m).

[0094] In some scenarios where ground reflection causes TOF depth difference anomaly, if the TOF depth difference ΔD is between 5-10cm, but the fluctuation is >3cm within 10 consecutive frames (0.167s), it is determined to be "reflection interference", which is not included in the raised area calculation. The TOF sensor integrates a polarizing filter, which can reduce the depth error caused by reflection interference to within ±3%.

[0095] In some scenarios of local light curtain failure, if a certain light curtain detection point is continuously interrupted (T 中断 >30s), but the adjacent detection points are not interrupted, and the TOF corresponding area has no protrusion (ΔD≤3cm), it is determined to be "light curtain failure", and a "light curtain single point failure alarm" is sent to the remote management platform, and it is not determined to be a residual object. In this way, the bottom light curtain adopts a "2+1" redundant design, each detection point corresponds to 2 emitting tubes + 1 receiving tube, and single element failure does not affect the detection result.

[0096] The step ensures "continuous stay" determination logic and "real object existence" at the same time, solves the problems of high false judgment rate and high missed judgment rate of traditional single sensor, and fully meets the design goal of "full scene, high precision and low power consumption" in the technical scheme. In addition, through the step, the response time of the residual object detection is less than 1s, the minimum volume of the identification is 5cm³, and the false judgment rate is less than 0.05%, which meets the identification requirements of small items such as certificates and keys.

[0097] As shown in step S30, Figure 1 According to the depth feature data source, the appearance feature data source collected by the dual-spectrum camera, the personnel trajectory state and the abnormal behavior determination result, the improved ReID algorithm is used to assign, mark and cancel the ID of the passing personnel; identity association across frames and sensors is realized, and trajectory tracing and abnormal behavior personnel positioning are supported.

[0098] Specifically, as Figure 4 The flowchart of step S30 provided by some embodiments is shown, and step S30 includes: extracting fusion features from the depth feature data source, the appearance feature data source, the personnel trajectory state and the abnormal behavior determination result (step S31); and assigning and associating the fusion features (step S32).

[0099] Specifically, step S31 includes:

[0100] Step S311: Obtain a background depth map from the depth feature data source, and remove personnel point clouds, wherein the personnel point clouds are point clouds with a depth difference of less than 5cm; use the farthest point sampling algorithm to sample 1024 key point clouds from the original point clouds, and retain the shape features of the personnel; align the personnel point clouds to a standard human coordinate system through T-Net transformation, and eliminate feature deviations caused by posture changes; perform 3-layer convolution operation on the sampled point clouds to extract local features; through the posture attention module, assign different weights to the features of the torso, head and limbs, and strengthen the feature contribution of stable regions; output a 32-dimensional depth feature vector f 深度 , and normalize the features while outputting feature confidence;

[0101] The depth feature data source is a TOF sensor array (frame rate ≥ 60 fps, depth measurement error ± 2%, point cloud density ≥ 100 points / cm2) deployed in the channel behavior judgment layer, which can collect the three-dimensional depth information of the personnel in real time. The depth feature data source includes depth point cloud data and dynamic contour data. The depth point cloud data is a three-dimensional coordinate point cloud of the whole body of the personnel (including the head, torso and limbs), and each frame contains ≥ 10000 valid points (excluding background noise points). The dynamic contour data is a sequence of 10 consecutive frames (about 0.167s) of depth point cloud, which is used to extract the dynamic morphological features of the personnel (such as shoulder width change and gait contour).

[0102] Step S312: Obtain visible light image and infrared image from appearance feature data source, and detect personnel target frame from visible light image and infrared image respectively by using YOLOv5-L network; use adaptive histogram equalization to improve the texture contrast of visible light image; use Gaussian filter to extract edge features of body contour from infrared image; correct image distortion based on TOF depth data, and crop the pixels of visible light image with improved texture contrast and the pixels of filtered infrared image to correspond to the spatial position of depth feature; use improved ResNet50 network, add a dual-modal fusion module to the 3rd and 4th layers of the network, input visible light image and infrared image into the first 2 layers of ResNet50 respectively to extract basic features, and fuse dual-modal features through attention weight at the 3rd layer; remove the last fully connected layer of the network, retain 2048-dimensional convolutional features, reduce them to 64-dimensional through PCA, and obtain appearance feature vector f RGB and fusion weight;

[0103] Step S313: Obtain the trajectory state of the personnel from the personnel trajectory state, extract the position data of multiple consecutive frames (about 0.33s), calculate the step frequency, step length and step width; extract the gait cycle fluctuation coefficient, torso swing angle, foot landing speed, left-right step length difference, left-right step frequency difference, step length to height ratio, step frequency to age correlation coefficient, swing phase time proportion, support phase time proportion, double foot off ground time, vertical displacement fluctuation, horizontal acceleration mean value and vertical acceleration mean value; standardize the step frequency, step length, step width, gait cycle fluctuation coefficient, torso swing angle, foot landing speed, left-right step length difference, left-right step frequency difference, step length to height ratio, step frequency to age correlation coefficient, swing phase time proportion, support phase time proportion, double foot off ground time, vertical displacement fluctuation, horizontal acceleration mean value and vertical acceleration mean value to obtain a 16-dimensional original gait vector; perform nonlinear transformation on the original vector through a 1-layer fully connected network (activation function ReLU) to obtain gait feature vector f 步态 and gait stability score.

[0104] Step S314: Sequentially concatenate the 32-dimensional depth features, 64-dimensional appearance features, and 16-dimensional gait features to obtain a 112-dimensional original fused feature vector; normalize the original fused feature vector; construct a two-layer fully connected multilayer perceptron network. The first layer is a hidden layer with 112-dimensional input and 128-dimensional output, using LeakyReLU as the activation function and adding a BatchNorm layer. The second layer is the output layer with 128-dimensional input and 256-dimensional output, without an activation function, outputting the fused feature vector f. 融合 A feature attention module is added to the fully connected multilayer perceptron network to calculate the attention weight of each original feature dimension. The 112-dimensional original vector is weighted and then input into the MLP network. The output is a 256-dimensional fused feature vector and the attention weights of each single-modal feature.

[0105] In this step, the personnel trajectory status X^ of the passage obtained in step S10 is used as the basis. t∣t The positional data of 20 consecutive frames (approximately 0.33 seconds) were extracted to calculate the following basic gait parameters:

[0106] Cadence frequency (cFM) refers to the number of footsteps landing per unit of time, measured in steps per second. This system calculates cFM by counting the number of complete gait cycles within 20 frames. For example, if there are two cycles within 20 frames (approximately 0.33 seconds), the cFM is f. 步频 =2 / 0.33≈6 steps / second. The system identifies the start and end points of gait cycles based on changes in the pedestrian's foot position to count the number of cycles. Step length refers to the distance the pedestrian moves horizontally (x-axis) within a gait cycle. The system records the pedestrian's horizontal coordinates at the start and end of each cycle, calculates the difference in step length, and averages the step lengths of multiple cycles to obtain the final step length L. 步长 Stride width refers to the average vertical distance (y-axis) between the two feet in each frame of an image during walking. The system identifies the position of the two feet in each frame and measures the vertical distance; the average of the measurements from multiple frames yields the stride width W. 步宽 The gait cycle fluctuation coefficient measures gait cycle stability and is calculated by the ratio of the standard deviation of multiple gait cycle durations to the average duration. A coefficient ≤ 5% indicates a stable gait. Trunk swing angle: The angle of trunk swing relative to the vertical direction during walking. The system analyzes TOF sensor depth point cloud data, calculates the swing angle based on changes in the trunk's central axis, and averages the angle over multiple time periods. An angle ≤ 15° is considered normal. Foot landing speed: The speed at which the foot contacts the ground. The system analyzes changes in foot position and time intervals to calculate this. Normal: speed ≤ 0.5 m / s. Left-right step length difference: the average of the difference between the left and right foot step length in each gait cycle, reflecting the difference between the left and right foot step length. Left-right step frequency difference: the difference between the left and right foot step frequency per unit time, evaluating the consistency of the left and right foot step frequency. Step length to height ratio: the ratio of the average step length to the estimated height, assisting in determining whether the walking pattern and body characteristics are consistent. Step frequency to age correlation coefficient: a model is built by statistical analysis of a large number of step frequency data of pedestrians of different ages. The estimated age of the current pedestrian and the theoretical step frequency calculated by the model are compared to obtain the coefficient, evaluating whether the step frequency is consistent with the age. Swing phase time ratio: the ratio of the swing phase time to the gait cycle time, calculated by the system to identify the start and end time of the swing phase. Support phase time ratio: the ratio of the support phase time to the gait cycle time. Double-foot off-ground time: the total time of double-foot off-ground in the gait cycle, calculated by the system to monitor the contact state and reflect the walking dynamic characteristics. Vertical displacement fluctuation: the vertical displacement fluctuation of the body center of gravity during walking, calculated by the system to record the vertical coordinate change and calculate the standard deviation as an indicator, reflecting the body stability. Horizontal acceleration average: the acceleration calculated from the horizontal position change of the pedestrian, averaged from multiple frames of acceleration, reflecting the horizontal acceleration trend. Vertical acceleration average: the acceleration calculated from the vertical position change of the pedestrian, averaged from multiple frames of acceleration, reflecting the vertical direction motion change.

[0107] Specifically, step S32 comprises:

[0108] Obtain the temporary trajectory ID from the personnel trajectory state, query the historical ID feature library, and the historical ID feature library is a fusion feature database storing allocated IDs;

[0109] Calculate the cosine similarity between the fusion feature of the current personnel and all features in the historical ID feature library for the fusion feature of the current personnel, for the historical fusion feature of the historical ID feature library; the cosine similarity is 0-1, and the closer to 1, the higher the matching degree;

[0110] If the cosine similarity is detected to be less than the cosine similarity threshold, it is determined to be a new personnel, a unique new ID is assigned to it, and the new ID is bound with its fusion feature, temporary trajectory ID, and first appearance timestamp and stored in the historical ID feature library, and the behavior state is synchronously associated with the mark "normal";

[0111] Obtain the fusion feature of each frame from the personnel trajectory state, and calculate the cosine similarity between the fusion feature and the corresponding ID feature in the historical ID feature library;

[0112] If the cosine similarity is detected to be greater than the cosine similarity threshold, update the feature of the ID to a sliding average feature; the cosine similarity threshold can be 0.85; the sliding average feature f 融合,更新 = 0.7⋅f 融合,历史 + 0.3⋅f融合,t Weighted averaging can smooth out feature fluctuations, such as feature changes caused by brief passenger occlusion.

[0113] If the detected cosine similarity is not greater than the cosine similarity threshold, the position distance between the current trajectory prediction position and the historical ID trajectory is calculated. If the position distance is less than or equal to 10cm, the original ID is retained and marked "feature to be recovered".

[0114] Output a unique new ID for each person, an updated historical ID feature database, and a table showing the association between IDs and historical personnel trajectories or behaviors;

[0115] Based on the status of personnel trajectory, when a passenger trajectory is detected to be outside the channel detection range and there is no feature match for 10 consecutive frames (≈0.167s), if the personnel behavior status is still detected to be normal, the ID is immediately cancelled and its fused feature data is deleted from the historical ID feature library; if the personnel behavior status is still detected to be abnormal, the ID is marked as pending archiving.

[0116] If an ID tag to be archived is detected and associated with complete motion trajectory status data, timestamp of abnormal behavior, type of abnormal behavior, on-site image, fused feature vector and single-modal feature data, then an abnormal ID archive file is generated.

[0117] In this step, TOF depth features are incorporated, unaffected by light conditions, allowing for the extraction of passengers' 3D morphological features (such as shoulder width and height ratio), retaining over 50% of effective features even in occluded scenarios. The addition of infrared thermal imaging features improves recognition accuracy to over 95% in nighttime scenes. An attention module automatically adjusts the weights of each feature based on the scene (e.g., 0.6 weight for infrared features in low-light scenes, and 0.5 weight for RGB features in normal scenes). This step demonstrates that after multimodal fusion, ID matching accuracy is >95%, and accuracy in occluded scenarios is >85%, representing a 25% improvement over traditional methods.

[0118] like Figure 1 As shown, the control method for the anti-backflow channel provided in the embodiment further includes step S40: calculating the channel passage width and the required number of channels. Specifically, step S40 includes:

[0119] Step S41: Determine the one-way traffic width and two-way traffic width based on the minimum shoulder width, turning space, and safety distance required for one-way pedestrian passage. 单通道 =R 转弯 +L 行人 +S 安全 W 双通道 =2×(R 转弯 +L 行人 +S 安全 ), where L 行人 R is the minimum shoulder width required for one-way pedestrian traffic.转弯 W 安全 is the safety distance, the coefficient "2" indicates the requirement of covering the simultaneous passage of two-way pedestrians, ensuring that there is no collision when two people pass each other; in some embodiments, W 通道(双向) = 2 x (0.5 + 0.6 + 2) = 2.6 m.

[0120] Step S42: predicting the current crowd density according to the weighted historical time series data, the prediction model being , t n is the previous n time slices (each slice is 10 s), a n is the correlation coefficient, is the predicted crowd density of the current frame, is the measured crowd density of the previous n frames; further, if the predicted crowd density is greater than 10% than the measured crowd density, the correlation coefficient is adjusted, for example, a1 is increased and a5 is decreased;

[0121] Step S43: determining the single-lane traffic capacity according to the average walking speed of pedestrians, the safe crowd density, and the one-way passage width, and determining the double-lane traffic capacity according to the average walking speed of pedestrians, the safe crowd density, and the two-way passage width;

[0122] C 单通道 = v 步行 x p 临界 x W 单通道， C 双通道 = v 步行 x p 临界 x W 双通道 ; v 步行 is the average walking speed of pedestrians, for example, v 步行 = 1.2 m / s (measured value, covering normal walking and slow queuing scenarios), p 临界 = 2 people / m 2 is the safe crowd density threshold (to avoid crowding and trampling).

[0123] Step S44: determining the required number of lanes according to the predicted crowd density of the current frame, the average walking speed of pedestrians, the one-way passage width, and the single-lane traffic capacity; during peak hours (such as the peak of airport flight arrivals), the lane utilization rate is increased by 30%, during flat peak hours (such as early morning), redundant lanes are closed, and the daily energy consumption of a single lane is reduced by 40%;

[0124] The number of single lanes N = , wherein is the ceiling function.

[0125] For example, Figure 1As shown, the control method of the anti-reflux passage provided by the embodiment further includes step S50: based on at least one of the passage passing width, the required passage number, and the abnormal behavior determination result, generating at least one of the traffic mode instruction, the free mode instruction, the interlock mode instruction, the lock mode instruction, the maintenance mode switching instruction, the emergency mode switching instruction, the high-risk abnormality corresponding instruction, the medium-risk abnormality corresponding instruction, and the low-risk abnormality corresponding instruction;

[0126] Specifically, step S50 includes:

[0127] Step S51: if it is detected that the current crowd density is less than 1 person / m 2 and the required passage number is 1, a traffic mode instruction is generated, the traffic mode instruction including a first door control instruction, a first sensor control instruction, and a first alarm control instruction, the first door control instruction being used to control the high-security level and low-security level doors in at least one passage to remain in an open state, the first sensor control instruction being used to control the sensors to be reduced to a low-power state, and the first alarm control instruction being used to control the reverse behavior alarm, the throwing behavior alarm, and the non-triggering sound and light alarm;

[0128] Step S52: if it is detected that the current crowd density is between 1 and 2 persons / m 2 and the required passage number is 1, a free mode instruction is generated, the free mode instruction including a second door control instruction, a second sensor control instruction, and a second prompt instruction, the second door control instruction being used to control the high-security level and low-security level doors in at least some passages to remain in a sensing state, the first sensor control instruction being used to control the TOF sensor to keep tracking the position of the person, and the second prompt instruction being used to control the passage prompt interface to remain in a display state of “normal and orderly passage”. The sensing state refers to that the person is within 0.8 m (based on the calibrated passage width of 2.8 m, the sensing area covers the full width of the passage) from the door body, the door body motor is started, opened within 0.3 s, and closed after the person leaves with a delay of 1 s. If the sensing area of the door body is too narrow (only covers a width of 1 m), the installation angle of the infrared probe needs to be adjusted (tilted outward by 15°) to ensure that the 2.8 m passage width is covered.

[0129] Step S53: if it is detected that the current crowd density is not less than 2 persons / m 2If the required number of channels is not less than 2, an interlocking mode instruction is generated, which includes a third gate control instruction and a third passage instruction, the third gate control instruction is used to control the high security level and low security level gate in all channels to keep only one open and control the flow of people in the channel content, the third prompt instruction is used to control the passage prompt interface to keep displaying the state of "current congestion, please queue up" and issue a low frequency buzzer alarm. Wherein, the high security level and low security level gate only one door keeps open is A door open → personnel enter → A door closed (confirm that the personnel enter the channel completely) → B door open → personnel leave → B door closed, every 2s release 1 person; When there are multiple channels, it is necessary to ensure that the action of each group of door bodies is not in conflict (such as after the A door of channel 1 is closed, the A door of channel 2 can be opened); And control the flow of people in the channel content is: TOF sensor counts the number of people in each channel, if the number of people in a channel is > 3 (based on 2.8m width × 7.5m length, safe capacity 6 people), send "pause release" instruction to the previous channel;

[0130] Step S54: If it is detected that the current flow density is zero and the required number of channels is zero, a lock mode instruction is generated, which includes a fourth gate control instruction, a fourth sensor instruction and a fourth security linkage instruction; The fourth gate control instruction is used to control the high security level and low security level gate in all channels to be locked, the door body controller cuts off the motor power supply, only the emergency key unlocking interface is reserved; The fourth sensor instruction is used to control the sensor to be reduced to an ultra-low power consumption state, and the fourth security linkage instruction is used to control whether there is an object moving in the detection channel and send an intrusion alarm according to the detection result; Wherein, the ultra-low power consumption state is that the TOF sensor is reduced to ultra-low power consumption (frame rate 10fps), and the light curtain only retains the "obstruction > 10s" alarm (to prevent malicious obstruction);

[0131] Step S55: According to the received maintenance mode switching instruction, generate the fifth door control instruction, the fifth alarm instruction and the fifth permission verification instruction. The fifth door control instruction is used to control the low security level door in the channel to remain closed, the high security level door to remain open and send the automatic closing prohibition instruction at the same time. The fifth alarm instruction is used to close the abnormal alarm function of the channel. The fifth permission verification instruction is used to identify whether the personnel is a maintenance personnel and determine whether to open the high security level door according to the identification result. This mode is a hardware maintenance mode. The remote management platform or the local maintenance terminal sends the instruction to forcibly switch the mode, and the priority is higher than that of the automatic mode. For example, replace the light curtain sensor and calibrate the door body. After the fifth door control instruction is generated, the B door of the maintenance channel is closed (the electromagnetic lock is powered on to block the external personnel from entering), and the A door remains open (to facilitate the maintenance personnel to enter and exit), and the “automatic closing prohibition” instruction is sent to the door controller at the same time. After the fifth alarm instruction is generated, the abnormal alarm function (such as the reverse and the left-over object alarm) of the channel is closed to avoid the maintenance action from being triggered by mistake. After the fifth permission verification instruction is generated, when the personnel wearing the authorized RFID card (the card number matches the system whitelist) approaches the A door, the pressure sensor allows the A door to be manually pushed open (otherwise, the A door remains half-locked).

[0132] Step S56: According to the received emergency mode switching instruction, generate the sixth door control instruction, the sixth alarm instruction and the sixth sensing instruction. The sixth door control instruction is used to control the high security level door and the low security level door in the channel to be opened at the same time or to be manually opened. The sixth alarm instruction is used to produce the sound and light alarm for emergency evacuation. The sixth sensing instruction is used to control the suspension of abnormal detection. This mode is an emergency mode, such as fire, equipment failure, etc. After the sixth door control instruction is generated, all A doors / B doors of the channel are immediately opened (the electromagnetic lock is powered off, and the door body can be manually pushed open. Even if the main power is interrupted, the standby power can also maintain the door body to be opened for ≥4 hours). The sound and light alarm can be “green light + circulating voice ‘please pass quickly, please do not stay’” (the volume is adjusted to the maximum 80 dB, covering the whole channel). After the sixth sensing instruction is generated, all abnormal detection (such as reverse and object throwing) is suspended, and only the personnel counting function of the TOF sensor is reserved to push the “evacuation person number real-time statistics” to the remote platform.

[0133] Step S57: If the continuous reverse behavior and the object throwing behavior are detected, generate the high-risk abnormal corresponding instruction. The high-risk abnormal corresponding instruction includes the seventh door control instruction, the seventh alarm instruction and the seventh linkage instruction. The seventh door control instruction is used to lock the high security level door and the low security level door in the abnormal channel. The seventh alarm instruction is used to produce the sound and light alarm of the high-risk abnormality. The seventh linkage instruction is used to send the abnormal coordinates and the abnormal real-time video stream.

[0134] For the seventh gate control instruction, after the door body controller receives it, the electromagnetic lock is powered on, and the response time is <0.3 s. If the pressure sensor of the door body detects a >50N impact force (person forcibly pushing the door), the automatic switching to the "half-lock state" (the door body retains a 5cm gap to avoid injury to the person) is performed. If the abnormal channel is multiple groups (N≥2), only the abnormal channel is locked, and the other channels operate normally to avoid global congestion; for the seventh alarm instruction, after the audible and visual alarm receives it, it emits "red light + high-frequency buzzing (2 times / sec)", which lasts until the anomaly is resolved, and at the same time, it pushes a "high-risk abnormal alarm" to the remote platform, accompanied by a 10s video clip; the channel entrance LED screen displays "abnormal area, do not approach", and marks the abnormal coordinates (such as "4.2m place, please detour"); for the seventh linkage instruction, after the security terminal receives it, the channel camera focuses on the abnormal area (zooms to 2 times to clearly capture the features of the person); if it is a throwing, the security terminal sends a "standby instruction" to the cleaning robot (do not go there temporarily to avoid interfering with security disposal), and executes the cleaning after the security confirms that there is no danger.

[0135] Step S58: If a single reverse behavior or a single wandering behavior is detected, a medium-risk abnormal response instruction is generated, which includes an eighth gate control instruction, an eighth alarm instruction, and an eighth guidance instruction. The eighth gate control instruction is used to control the high-security level door to delay closing and the low-security level door to normally close according to the single reverse behavior, and to control the door body to enter the free mode according to the single wandering behavior; the eighth alarm instruction is used to generate an audible and visual alarm for a medium-risk abnormality, and the eighth guidance instruction is used to issue prompt display information to guide the person passing through who has the single reverse behavior or the single wandering behavior;

[0136] For the eighth gate control instruction, only A door is controlled to delay closing for 2s (B door remains in the normal state), to avoid misjudgment leading to interruption of passing, and if the person resumes normal passing within 2s, the door body normally closes; if it is wandering, the door body remains in the current mode (such as free mode), but sends a "speed reduction prompt" instruction to the door controller, to extend the door opening delay to 2s, to guide the person to pass quickly; for the eighth alarm instruction, after the audible and visual alarm receives it, it emits an alarm of "yellow light + low-frequency buzzing (1 time / sec)" for 3s and then automatically stops, to avoid disturbing the public; at the same time, it sends a "medium-risk abnormality record" to the remote platform, accompanied by the abnormal person ID and the trajectory clip, without triggering a real-time pop-up window (only stored in the log); for the eighth guidance instruction, after the LED screen in the channel receives it, it displays "please pass in the indicated direction" or "please do not stay, pass quickly", and if the wandering person continues for >10s, it pushes a prompt to the APP of the nearby patrol security (non-mandatory disposal).

[0137] Step S59: If the left-over is detected, a low-risk abnormal response instruction is generated, including a ninth gate control instruction, a ninth alarm instruction, and a ninth linkage instruction. The ninth gate control instruction is used to control the door body to enter the flow mode or the free mode. The ninth alarm instruction is used to generate an audible and visual alarm for the left-over. The ninth linkage instruction is used to link the cleaning robot to clean the left-over.

[0138] For the ninth gate control instruction, the door body remains in the current mode (such as the flow mode or the free mode) and does not interrupt the passage, but the light curtain near the left-over position triggers the "slow prompt". When a person is detected to approach, a "delay door opening for 0.5s" instruction is sent to the door controller to avoid the person from kicking the left-over. For the ninth alarm instruction, the audible and visual alarm sends out "blue light + voice prompt 'there is a left-over in the passage / abnormal equipment, please handle in time'" (volume 60dB, to avoid interfering with normal passage). A "low-risk abnormal work order" is sent to the remote platform, and the disposal priority is marked (left-over > slight failure). For the ninth linkage instruction, the cleaning robot receives the abnormal coordinates, and the robot goes to the left-over according to the "shortest path". After arriving at the left-over, a "arrival notification" is sent. After cleaning is completed, a "disposal completion" is fed back to the platform.

[0139] The above only describes the embodiments of the present application and is not configured to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the scope of the claims of the present application.

Claims

1. A method of controlling a reflux-preventing passage, characterized by, The method comprises the following steps: According to the obtained personnel depth map centroid coordinates, object depth map centroid coordinates, and penetration point coordinates, and the historical trajectory state, the updated personnel trajectory state, object trajectory state, and trajectory association result are obtained, the trajectory association result includes at least one of association success, new trajectory generation, and trajectory termination; Based on the personnel trajectory state, object trajectory state, and trajectory association result, ReID feature, acquired light curtain timing data, and acquired camera pose data, the reverse behavior, wandering behavior, throwing behavior, and abandoned object in the channel are determined to obtain an abnormal behavior determination result; According to the depth feature data source, the appearance feature data source collected by the dual-spectrum camera, the personnel trajectory state, and the abnormal behavior determination result, the improved ReID algorithm is used to assign, mark, and cancel the ID of the passing personnel; The channel passing width and the required channel number are calculated; Based on at least one of the channel passing width, the required channel number, and the abnormal behavior determination result, at least one of the flow mode instruction, the free mode instruction, the interlocking mode instruction, the locking mode instruction, the maintenance mode switching instruction, the emergency mode switching instruction, the high-risk abnormal corresponding instruction, the medium-risk abnormal corresponding instruction, and the low-risk abnormal corresponding instruction is generated.

2. The control method according to claim 1, characterized by, According to the obtained personnel depth map centroid coordinates, object depth map centroid coordinates, and penetration point coordinates, and the historical trajectory state, the updated personnel trajectory state, object trajectory state, and trajectory association result are obtained, the trajectory association result includes at least one of association success, new trajectory generation, and trajectory termination, comprising: According to the collected personnel depth map centroid coordinates, object depth map centroid coordinates, and penetration point coordinates, the data fusion is performed to obtain the fusion observation value vector and the observation value vector of the personnel; Based on the fusion observation value of the current frame, the historical trajectory state, and the ReID feature similarity, the trajectory association result is obtained, the trajectory association result includes at least one of the association success result, the generated new trajectory, the updated trajectory, and the trajectory termination result; and The trajectory termination is judged.

3. The control method according to claim 2, characterized by, According to the collected personnel depth map centroid coordinates, object depth map centroid coordinates, and penetration point coordinates, the data fusion is performed to obtain the fusion observation value vector and the observation value vector of the personnel, comprising: The personnel depth map centroid coordinates, object depth map centroid coordinates, and penetration point coordinates with the same sampling frequency are obtained; The personnel motion state vector or the object motion state vector is determined according to the personnel depth map centroid coordinates, the object depth map centroid coordinates and the penetration point coordinates, and the motion state vector is X t = [x, y, v x ,v y ,a x ,a y ] T , wherein x and y are two-dimensional positions, v x ,v y are horizontal and vertical velocities respectively, and a x ,a y are horizontal and vertical accelerations respectively. establishing a discrete-time state equation where X t is a motion state vector for a t frame, X t-1 is a motion state vector for a t-1 frame, F is a state transition matrix, is a process noise for a motion disturbance in a time interval from a t-1 frame to a t frame; The observation equation is established as Z t = HX t + V t , wherein Z t is an observation value vector of a t frame, X t is a motion state vector of the t frame, H is an observation matrix, and V t is an observation noise vector. According to the discrete-time state equation and the observation equation, a person observation value vector or an object observation value vector is obtained, and the observation value vector is Z t =[Z t , x ,Z t,y ] T , wherein Z t,x =ω tof x tof +ω light x light , Z t,y =ω tof y tof +ω light y light , Z t , x ,Z t,y is a two-dimensional position coordinate of the person or the object in the channel plane determined according to the observation result, ω tof is a depth map centroid coordinate weight, and ω light is a penetration point coordinate weight.

4. The control method according to claim 2, characterized by, Based on the fusion observation value of the current frame, the historical trajectory state, and the ReID feature similarity, the trajectory association result is obtained, the trajectory association result includes at least one of the association success result, the generated new trajectory, the updated trajectory, and the trajectory termination result, comprising: The historical trajectory state is obtained according to Kalman prediction, and the historical trajectory state is filtered estimation value of a t-1 frame and a covariance matrix , a prior covariance The coordinates in the prior covariance t│t−1 quantify the uncertainty of the predicted position of the historical trajectory in the current frame The personnel recognition feature is extracted from the collected image, a ReID feature vector is acquired, the ReID feature vector includes current frame features and historical track features, and ReID feature similarity is determined according to the current frame features and the historical track features; the ReID feature similarity S ReID,ij =f t f t-1 / ║f t ║║f t-1 ║,f t ,f t-1 are respectively current frame features and historical track features. The association cost of the current observation target and the historical trajectory is calculated; It is determined whether to match the observation target of the current frame with the historical trajectory according to the association cost; If it is detected that the association cost of a certain observation target is less than the cost matrix threshold value, and the association cost is the minimum cost of the observation target and the minimum cost of the historical trajectory, it is determined that the association is successful; If it is detected that the association cost of multiple historical trajectories and the same observation target is less than the cost matrix threshold value, the historical trajectory corresponding to the minimum association cost is selected as the association result; If all the association costs of a certain observation target are detected to be not less than the cost matrix threshold, and the number of observation targets in the current frame is greater than the number of historical trajectories, it is determined that a new trajectory, and the trajectory state of the target is initialized and a new trajectory ID is assigned; If the prior covariance of the historical trajectory is detected If the covariance exceeds a preset threshold, a step of judging the consistency of the moving direction of the current observed target and the velocity vector direction of the historical trajectory is performed. If the association is detected to be successful, the posterior covariance P obtained by Kalman update t∣t Still exceeds the covariance preset threshold, mark the track as "to be confirmed association".

5. The control method according to claim 2, characterized by, The trajectory termination is judged, including: Recording the association state of the trajectory; Judging the timing interruption and the occlusion timeout; Determining whether to trigger the trajectory termination according to the coordinates of the spatial position; Performing the trajectory termination operation.

6. The control method according to claim 1, characterized by Based on the personnel trajectory state, the object trajectory state, the trajectory association result, the ReID feature, the obtained light curtain timing data and the obtained camera pose data, the reverse behavior, the loitering behavior, the throwing behavior and the left object in the channel are determined, and an abnormal behavior determination result is obtained, including: Determining the reverse behavior of personnel in the channel based on the consistency result of the personnel torso orientation and the movement direction, and the direction of the channel safety level; Quantifying the uncertainty of the movement direction by calculating the information entropy of the personnel trajectory direction angle, and identifying the loitering behavior without clear traffic intention; Determining the throwing behavior of personnel in the channel according to the captured object movement characteristics and personnel action characteristics; According to the comparison between the detection result of the continuous occlusion detection and the static profile, the left object is identified according to the comparison result.

7. The control method according to claim 6, characterized by Determining the reverse behavior of personnel in the channel based on the consistency result of the personnel torso orientation and the movement direction, and the direction of the channel safety level, including: Normalizing the speed vector direction in the trajectory state vector of the personnel, so as to match the torso orientation angle dimension; If the torso orientation angle is detected to be less than 30°, it is marked as the first candidate reverse behavior; if the absolute value of the angle difference between the speed vector direction and the torso orientation angle is less than 15°, it is marked as the second candidate reverse behavior; if the torso orientation angle is detected to be less than 30° and the absolute value of the angle difference between the speed vector direction and the torso orientation angle is greater than 15°, it is marked as the posture abnormal behavior; Maintaining some reverse candidate counters for each trajectory, when a certain frame meets the "reverse candidate" condition, the count is increased by 1; when a certain frame does not meet the condition, the counter is cleared; when the counter is greater than or equal to 3, the displacement verification step is entered; if a certain frame or invalid, the counter is suspended; Based on the displacement coordinates of the channel length direction in the trajectory association result, the reverse displacement amount is calculated, and if the reverse displacement amount is greater than the displacement threshold, it is marked as the third candidate reverse behavior; The reverse behavior is screened and verified from the first candidate reverse behavior, the second candidate reverse behavior and the third candidate reverse behavior.

8. The control method according to claim 6, characterized by, Quantifying the uncertainty of the movement direction by calculating the information entropy of the personnel trajectory direction angle, and identifying the loitering behavior without clear traffic intention, including: Discretizing the speed vector direction to obtain multiple direction angle intervals; a direction angle probability for each direction angle interval is calculated according to the formula , nk is the number of angles appearing in the kth interval, is the total number of valid angles in 10 frames, ≤ 10, and the sum of the direction angle probabilities of all direction angle intervals is equal to 1; Calculating the trajectory entropy according to the direction angle probability, and determining the loitering behavior according to the trajectory entropy, the stay time and the moving distance: If the trajectory entropy of the personnel in the channel is detected to be greater than 1.8, the stay time is greater than 5s and the moving distance is less than 1m, the behavior of the personnel in the channel is judged as loitering behavior, and the trajectory entropy, the stay time and the moving distance are recorded; If the trajectory entropy of the personnel in the channel is detected to be greater than 1.8, the stay time is greater than 5s and the moving distance is equal to 1.2m, the behavior of the personnel in the channel is marked as "to be re-determined"; If the trajectory entropy of the person in the channel is not greater than 1.8, the stay time is not greater than 5s, and the moving distance is not less than 1m, or one of the trajectory entropy of the person in the channel is not greater than 1.8, the stay time is not greater than 5s, and the moving distance is not less than 1m, the behavior of the person in the channel is determined as normal behavior, and the wandering determination counter is reset.

9. The control method according to claim 6, characterized by, According to the captured object motion characteristics and the person action characteristics, the throwing behavior of the person in the channel is determined, including: According to the real-time analysis of the time sequence data output by the light curtain sensor, it is detected whether there is an abnormal light shielding event, according to the timestamp of the light shielding event and the sensor position information, the motion trajectory of the object is calculated by using the least square method and other algorithms; From the skeleton point data, the posture features closely related to the throwing behavior are extracted, and the sliding window technology is used to analyze the posture data of continuous multiple frames to judge the trend of the posture change; The light curtain time sequence data and the human body posture data are time aligned to ensure that the two kinds of data at the same time point can be accurately associated, and the weighted fusion method is used to fuse the features obtained by the light curtain time sequence analysis and the features obtained by the human body posture analysis; If the light curtain detects an object crossing event, and the object moving speed exceeds 2m / s, or the time interval of the object crossing the light curtain is less than 0.3s, the event is marked as a suspicious throwing event; If the human body posture recognition result shows that the arm stretching angle exceeds 50, and the angle changes more than 30° within 0.5s, and the twist angle exceeds 20°, the posture is marked as a suspicious throwing posture; Check whether there is a suspicious throwing posture matching the time point corresponding to the suspicious throwing event; if the light curtain detects an object crossing event at the same time, the person's posture is in the throwing preparation or execution state, and both have consistency in time and space, the throwing behavior determination weight is increased; According to the throwing behavior confidence calculated by multi-source data fusion, the confidence is compared with the preset confidence threshold, if the confidence is greater than the threshold, the throwing behavior is determined.

10. A control system for an anti-regurgitation passageway, characterized by Including: The trajectory state tracking module is configured to obtain updated person trajectory state, object trajectory state and trajectory association result according to the obtained person depth map centroid coordinates, object depth map centroid coordinates and penetration point coordinates and historical trajectory state, the trajectory association result includes at least one of association success, new trajectory generation and trajectory termination; The abnormal behavior determination module is configured to determine the reverse behavior, wandering behavior, throwing behavior and left object in the channel based on the person trajectory state, object trajectory state and trajectory association result, ReID feature, obtained light curtain time sequence data and obtained camera posture data, and obtain abnormal behavior determination result; The ID module is configured to assign, mark and cancel the ID of the passing person through the improved ReID algorithm according to the depth feature data source, the appearance feature data source collected by the dual-spectrum camera, the person trajectory state and the abnormal behavior determination result; The channel calculation module is configured to calculate the channel passing width and the required number of channels; The instruction control module is configured to generate at least one of a flow mode instruction, a free mode instruction, an interlock mode instruction, a lock mode instruction, a maintenance mode switching instruction, an emergency mode switching instruction, a high-risk abnormality corresponding instruction, a medium-risk abnormality corresponding instruction, and a low-risk abnormality corresponding instruction based on at least one of the lane passage width, the required number of lanes, and the abnormality behavior determination result.

Images