A mixed reality school-entering and exiting management method and system based on visual intention prediction

Abstract

The application discloses a mixed reality access control method and system based on visual intention prediction, and relates to the technical field of intelligent campus security. Based on fixed-view camera continuous video frames, a human body passing intention state tensor aligned with a gate is constructed; according to the state tensor and an identity verification result, a visual intention risk score is generated by fusing behavior intention probability, structural anomaly score and identity constraint; based on the risk score, a spatially bound holographic early warning label is generated; according to the risk score and the holographic early warning label, gate control action control and mixed reality linkage output are executed, so that pre-passing risk early warning is realized, the rate of gate misopening and missed judgment is reduced, and the efficiency of on-site linkage disposal is improved.

Description

Technical Field

[0001] This invention belongs to the field of smart campus security technology, specifically relating to a mixed reality access control method and system based on visual intent prediction. Background Technology

[0002] As the construction of smart campuses continues to advance, campus entrances and exits have evolved from simple card-swiping scenarios to complex security scenarios integrating video sensing, identity verification, alarm linkage, and access control. These scenarios are characterized by high pedestrian traffic, rapid behavioral shifts, complex identity types, and short durations of pre-risk warning signs. Relying solely on static identity verification makes it difficult to identify risky individuals whose identity has been verified but whose behavior is abnormal; relying solely on access control opening / closing results or boundary violations makes it difficult to promptly detect pre-access risks such as tailgating, probing, reverse turning, and unauthorized entry, easily leading to accidental access control openings, delayed responses, and increased workload for manual verification.

[0003] In existing solutions, one type of method focuses on monitoring rules and risk classification for entrance and exit events. For example, Chinese patent CN202610750920.8 discloses an intelligent monitoring and early warning method for vehicle entry and exit on campus, which mainly focuses on graded early warning based on regional speed, dwell time, and risk range. This type of method is more suitable for regulatory tasks based on regional outcome states, but it lacks fine-grained modeling capabilities for personnel approach direction, passage preparation actions, and close-range tailgating relationships in campus gate scenarios. Therefore, it is difficult to complete pre-intervention before passage occurs, thereby increasing the risk of on-site misjudgment and missed judgment.

[0004] Another type of method focuses on video anomaly analysis and visitor behavior comparison. For example, Chinese patent CN202610365723.4 discloses a smart campus security management method and system based on video analysis, and Chinese patent CN202610052644.8 discloses a campus security control system integrating security monitoring and visitor management. These methods typically identify anomalies based on image sequences, historical trajectories, or behavioral baselines, improving post-event analysis capabilities. However, their risk assessment is often based on the fact that abnormal segments have already formed or trajectory deviations have accumulated. They lack a direct modeling mechanism for the sub-second intention formation stages such as forward probing, approaching, retreating, and tailgating preparations in entrance and exit scenarios, making it difficult to directly translate the identification results into control actions before access control is executed. This can lead to risky objects being discovered only after the gate action has been triggered during peak passage periods, increasing the risk of accidental gate openings and delays in security response.

[0005] Furthermore, most existing video security solutions perform detection, tracking, and alarming within a two-dimensional image pixel coordinate system. The pixel representation of the same action varies significantly under different fixed-view cameras, resulting in feature distribution being significantly affected by installation height, tilt angle, channel width, and degree of obstruction. This necessitates frequent parameter adjustments during deployment, increasing project implementation costs. Meanwhile, traditional alarms are typically delivered via background pop-ups, monitoring screen prompts, or audio-visual alerts. Monitoring personnel still need to perform a secondary matching between the monitored image and the on-site crowd, making it difficult to quickly identify specific risk targets and their locations, thus reducing on-site coordination efficiency. Summary of the Invention

[0006] The purpose of this invention is to provide a mixed reality access control method and system based on visual intent prediction, to solve the problems in existing technologies such as insufficient pre-intervention, accidental access control, and reduced on-site handling efficiency caused by the lack of continuous state modeling of the pre-access intent formation stage, the separation of identity verification and behavior discrimination links, and the difficulty in directly binding alarm results with the actual spatial location of the gate. To achieve the above objective, this invention provides a mixed reality access control method based on visual intent prediction, comprising: Step S1: Based on the continuous video frames captured by the calibrated fixed-view camera, construct the human passage intention state tensor aligned with the gate. Step S2: Based on the human body passage intention state tensor and identity verification results, generate a visual intention risk score that integrates identity and behavior; Step S3: Based on the visual intent risk score, generate a spatially bound holographic warning tag; Step S4: Based on the visual intent risk score and the holographic warning label, execute the access control action management and mixed reality linkage output.

[0007] Preferably, constructing the human passage intention state tensor for gate alignment includes: performing human pose extraction on the continuous video frames to obtain a two-dimensional skeletal keypoint sequence for each target within a continuous time window, and reading the gate reference endpoint and the gate physical width; establishing a gate alignment normalized coordinate system based on the gate reference endpoint with the gate's lateral and normal directions as basis vectors, mapping the two-dimensional skeletal keypoint sequence to the gate alignment normalized coordinate system, and converting it to metric local coordinates in conjunction with the gate physical width; and extracting the human centroid trajectory, propulsion speed, lateral contact speed, torso orientation consistency, ankle forward probing, ankle forward probing difference, and current centroid normal distance based on the metric local coordinates according to the midpoints of the shoulders, midpoints of the hips, and ankle keypoints to obtain a single-unit pose feature matrix.

[0008] Preferably, the construction of the gate-aligned human passage intention state tensor further includes: calculating the normal distance difference between adjacent targets based on the human centroid trajectory of each target in the individual posture feature matrix; constructing the lateral projection interval of each target based on the lateral coordinates of the shoulder key points and hip key points of each target in the metric local coordinates, and calculating the lateral projection overlap rate between adjacent targets based on the lateral projection interval; calculating the propulsion direction vector of each target based on the human centroid trajectory, and calculating the propulsion direction similarity between adjacent targets based on the propulsion direction vector of each target; calculating the trailing coupling strength based on the normal distance difference, lateral projection overlap rate, and propulsion direction similarity, combined with a preset distance attenuation coefficient, and selecting a preset number of neighboring targets with the highest coupling strength for each target to form a frame-by-frame trailing coupling vector; and concatenating the frame-by-frame trailing coupling vector with the individual posture feature matrix frame by frame to obtain the human passage intention state tensor.

[0009] Preferably, generating a visual intent risk score that integrates identity and behavior includes: receiving identity verification results returned by a card reader, a visitor appointment platform, and a faculty database, and encoding the identity verification results into a static identity constraint vector, which includes document matching status, passage time matching status, authorized channel matching status, and peer verification status; based on the identity verification time of the current target, searching for adjacent targets in the security zone in front of the same channel gate within a preset time window before and after, and determining the peer verification status based on the tail coupling strength, independent identity verification results, and similarity of the advancing direction of the adjacent targets; extending the static identity constraint vector to the time dimension and concatenating it with the human passage intent state tensor to obtain a fusion input tensor, and using the fusion input tensor to generate the visual intent risk score.

[0010] Preferably, the step of generating a visual intent risk score that integrates identity and behavior further includes: inputting the fusion input tensor into a temporal feature extraction layer to extract a hidden state sequence corresponding to a continuous time window; generating a short-window hidden state based on the hidden state sequence; inputting the short-window hidden state into an intent mapping layer to output an intent probability vector, wherein the intent probability vector corresponds to normal passage, probing and lingering, reverse reversal, tailgating and rushing through the gate, and unauthorized intrusion.

[0011] Preferably, the generation of the visual intent risk score that integrates identity and behavior further includes: obtaining the hidden state center of the normal passage sample and a preset scale parameter; calculating the distance between the short window hidden state and the hidden state center of the normal passage sample; and performing an exponential mapping based on the distance and the preset scale parameter to obtain a structural anomaly score.

[0012] Preferably, generating a visual intent risk score that integrates identity and behavior specifically includes: extracting the probability of tailgating and rushing through a gate, the probability of unauthorized intrusion, and the probability of reversing back from the intent probability vector to obtain high-risk behavior items; calculating identity mismatch items based on the matching of each component of the static identity constraint vector; multiplying the high-risk behavior items, structural anomaly scores, and identity mismatch items by their corresponding preset fusion weights, and summing the multiplication results to obtain the visual intent risk score.

[0013] Preferably, generating a spatially bound holographic warning tag includes: calculating the spatial anchor point of the holographic warning tag based on the human body's centroid trajectory, the propulsion direction vector, and the current centroid normal distance, combined with a preset lower limit of warning extrapolation distance, a preset upper limit of warning extrapolation distance, and a preset tag upward movement distance; generating a warning level index based on the visual intent risk score and the distance amplification relationship between the current centroid normal distance and a preset safety distance, combined with a preset distance amplification coefficient, and determining the risk level based on the warning level index; generating a holographic warning tag containing the object identity type, dominant risk intent, risk level, suggested action, and channel number based on the dominant risk intent, static identity constraint vector, and current channel number in the intent probability vector; and mapping the spatial anchor point to the terminal display coordinate system through the transformation relationship from the gate local coordinate system to the mixed reality terminal display coordinate system to obtain the spatially bound holographic warning tag.

[0014] Preferably, the execution of access control actions and mixed reality linkage output includes: comparing the visual intent risk score with preset release thresholds and preset interception thresholds, and generating release control instructions, delayed review instructions, or lock alarm instructions in combination with the identity verification results; when multiple targets exist in the same channel within the same processing cycle, priority is given to generating access control decisions based on the target closest to the gate passage line and with the highest visual intent risk score; sending the release control instructions, delayed review instructions, or lock alarm instructions to the access control execution interface and receiving access control receipts; when any of the following conditions are met: access control receipt timeout; no receipts received continuously; interface abnormality; then maintaining the gate closing priority degraded control state; sending the spatial anchor point, risk level, label content, and target bounding box coordinates of the holographic warning label to the mixed reality terminal, and sending them to the monitoring screen when the mixed reality terminal is not connected, completing the synchronous output of access control actions and mixed reality warnings.

[0015] This invention also provides a mixed reality access control system for schools based on visual intent prediction, comprising:

[0016] The state tensor construction module is used to construct a human passage intention state tensor aligned with the gate based on continuous video frames captured by a calibrated fixed-view camera; the risk score generation module is used to generate a visual intention risk score that integrates identity and behavior based on the human passage intention state tensor and the identity verification result.

[0017] The holographic tag generation module is used to generate spatially bound holographic warning tags based on the visual intent risk score;

[0018] The linkage control output module is used to implement the steps of the mixed reality access control method based on visual intent prediction as described in any one of claims 1 to 9, according to the visual intent risk score and the holographic warning label.

[0019] This invention has the following beneficial effects: By constructing a human passage intention state tensor aligned with the gate, the key points of the human skeleton, the advancing speed, the lateral contact speed, the consistency of the torso orientation, the foot and ankle forward probing difference, and the coupling relationship of multiple people following are uniformly mapped to a local coordinate system consistent with the physical space of the gate. This solves the problem in the prior art that it is difficult to express the intention formation process before passage by relying only on pixel displacement, dwell time or result state rules. As a result, the system can output risk judgment before the object passes through the gate, reducing the cost of false access and manual traceability caused by recognition lag.

[0020] By constructing a static identity constraint vector from document matching status, passage time matching status, authorized channel matching status, and peer verification status, and integrating it with behavioral intent prediction results and structural anomaly scores to form a visual intent risk score, the problem of the separation between the identity verification link and the behavior discrimination link in existing technologies is solved. This enables the simultaneous coverage of composite risk scenarios of "illegal identity and abnormal behavior" and "legal identity but abnormal behavior", reducing the false negative rate of high-risk events and improving the stability of risk identification in complex channel scenarios.

[0021] By further converting visual intent risk scores into spatially bound holographic warning tags and outputting them synchronously with access control commands, the problem of separation between traditional background pop-ups or ordinary duty screen alarms and on-site objects is solved. This allows duty personnel to directly see risk objects, risk levels, and suggested actions in the mixed reality terminal, shortening the cognitive link from alarm to action and improving the efficiency of linkage and handling and the continuous operation capability of the system in high real-time scenarios at entrances and exits. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is an overall flowchart of a mixed reality access control method and system based on visual intent prediction.

[0024] Figure 2 This is a flowchart of the human passage intention state construction unit in an embodiment of the present invention.

[0025] Figure 3 This is a flowchart of the risk score generation unit in an embodiment of the present invention.

[0026] Figure 4 This is a flowchart of the holographic early warning tag generation unit in an embodiment of the present invention.

[0027] Figure 5 This is a flowchart of the access control linkage output unit in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings.

[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0030] Example 1: Overall Process of a Mixed Reality School Access Control Method and System Based on Visual Intent Prediction

[0031] like Figures 1 to 5 As shown, this embodiment provides a mixed reality access control method and system based on visual intent prediction, executed by a mixed reality access control device. The mixed reality access control device includes a processor, a memory, a video access interface, an access control interface, and a mixed reality display interface. The processor, acting as the unified logic execution entity, receives continuous video frames uploaded from a fixed-view camera, identity verification results returned from a card reader and a visitor reservation platform, and outputs control results to the access control execution interface and the mixed reality terminal.

[0032] This embodiment uses a dual-channel scenario at a school main gate as an example. The main gate has two swing gate channels, each 0.9m wide, with a peak throughput of 48 to 65 people per minute. The camera frame rate is 25 fps, and the processing cycle is 0.2s. To ensure stable temporal characteristics are formed within a short time, the preset time window in this embodiment refers to a continuous frame window used to characterize the short-term attitude evolution before passage. This parameter can be determined jointly by the camera frame rate and the gate response delay. In practical engineering applications, its typical value is 5 frames, corresponding to... Furthermore, the configuration can typically be reasonably set between 3 and 8 frames in different scenarios. Seventeen skeletal keypoints are extracted for each target. To characterize the nearest neighbor following relationship, the preset number of neighboring targets in this embodiment refers to the number of highly relevant neighboring objects retained per frame for participating in the following relationship modeling. This parameter can be obtained through offline statistics from peak pedestrian flow video recordings, and its typical value is 2. In practical engineering applications, it can usually be configured between 1 and 4.

[0033] Step S1: Based on the continuous video frames captured by the calibrated fixed-view camera, construct the human passage intention state tensor aligned with the gate.

[0034] In this embodiment, this step is performed by the processor. Figure 2 The process shown is completed to uniformly map human postures from different camera positions to the passage status representation in the physical space scale of the turnstile, so that subsequent risk assessment can be carried out directly around the safe area in front of the turnstile.

[0035] Step S101: Generate human-gate normalized posture coordinates

[0036] Human pose extraction is performed on the continuous video frames to obtain the two-dimensional skeletal key point sequence of each target within a continuous time window, and the gate reference endpoint and gate physical width are read; based on the gate reference endpoint, a gate alignment normalized coordinate system is established with the gate's horizontal and normal directions as base vectors, the two-dimensional skeletal key point sequence is mapped to the gate alignment normalized coordinate system, and combined with the gate physical width to convert it into metric local coordinates.

[0037] In this embodiment, the processor first reads the last 5 frames of video from the video access interface and calls the human pose extraction program to obtain the two-dimensional skeletal key points of each target in each frame. The human pose extraction program can be implemented using an existing two-dimensional pose estimation model, as long as it can stably output key points such as the shoulders, hips, and ankles. Subsequently, the processor reads the gate reference endpoints and the gate physical width from the channel configuration file. Here, the engineering meaning of the gate reference endpoints is the calibration points at both ends of the gate passage line in the image, which can be manually calibrated through ground markers during the installation and commissioning phase; in this embodiment, the actual typical configuration corresponding to the aforementioned gate reference endpoints is as follows: , The physical width of the gate is used to restore the normalized coordinates to the actual length scale, and can be obtained by verifying the consistency between the width on the equipment nameplate and the on-site remeasurement value; in this embodiment, its typical value is 0.9m, and it is usually between 0.6m and 1.2m in similar campus swing gates.

[0038] To achieve the mapping from camera pixel coordinates to gate-aligned normalized coordinates, the processor constructs local basis vectors along the gate's horizontal and normal directions, using the gate's center as a reference, and then... The first goal in Frame number The pixel coordinates of each key point are normalized. This formula corresponds to the coordinate unification processing in this step. The input data are the pixel coordinates of the key points, the reference endpoint of the gate, and the gate width calibration result. It is used to solve the problem of lack of uniform spatial scale for the same action under different viewpoints. The output result is the gate-aligned normalized coordinates. The execution subject is the processor, and the timing of the call is after the human pose extraction in each processing cycle.

[0039] ,in, Indicates the first The first goal in Frame number Align the gates at key points with normalized coordinates; Represents the pixel coordinates of the original image, in pixels; This indicates the pixel length corresponding to the reference width of the gate; Indicates the center point of the turnstile; Represents a unit vector along the transverse direction of the turnstile; This represents the normal unit vector perpendicular to the gate's passage line and pointing into the campus. This transformation is based on two-dimensional geometric projection relationships and is applicable when the camera is fixed and the gate's reference endpoint has been installed and calibrated. When... If the value is abnormally small and may cause instability, the processor will not continue to use the current calibration result, but will instead call the most recent valid calibration value.

[0040] After normalization, the processor further recovers the metric local coordinates using the physical width of the gate. This relationship is used to ensure that subsequent calculations of velocity, distance, and spatial anchor points have a unified dimension.

[0041] ,in, This is a metric local coordinate system, with both components in meters. This refers to the physical width of the turnstile. During online operation, the processor prioritizes using the field-tested value from the channel configuration file; if this value is missing, it calls the factory default width and records this calculation as pending verification and calibration.

[0042] In this embodiment, the preset minimum width threshold is used to determine whether the gate width pixel calibration has failed. It can be obtained by statistically analyzing the lower limit of historical calibration results during the installation and debugging phase. The typical physical width of the gate is 0.30m, and the pixel threshold corresponding to the gate reference width is calculated accordingly. If the current calibration width is lower than this threshold, the processor directly falls back to the most recent valid calibration result. For cases where key points are missing, if some key points in the current frame are unavailable, the gate alignment normalized coordinates of the corresponding key points in the previous frame of the same target are used to fill in the missing points; if the missing points are in the first frame, the linear interpolation results of the shoulder and hip points on the same side are used to fill in the missing points. After this processing, video pose data from different sources are mapped to a unified local gate scale, and subsequent models can directly make judgments around the real physical space of the channel.

[0043] In the actual processing of this embodiment, the pixel coordinates of the midpoint of the shoulders of target A in the 5th frame are: The pixel coordinates of the midpoint of the two hips are After the aforementioned mapping, the normalized lateral component of the centroid of target A in frame 5 is 0.04, and the normal component is 1.06. Further conversion to metric local coordinates yields a lateral component of approximately 0.036m and a normal component of approximately 0.954m, indicating that the target is located slightly to the right of the gate's center and approximately 0.95m from the gate. The above mapping results constitute the aforementioned metric local coordinates.

[0044] Step S102: Extract propulsion, orientation, and lateral attachment features.

[0045] Based on the metric local coordinates, the trajectory of the human body's center of mass, propulsion speed, lateral contact speed, torso orientation consistency, foot and ankle forward probing, foot and ankle forward probing difference, and current center of mass normal distance are extracted according to the midpoints of the shoulders, the midpoints of the hips, and the key points of the feet and ankles to obtain the individual posture feature matrix.

[0046] In practical engineering applications, this step involves the following processing steps: The processor calculates the midpoints of the shoulders and hips based on the left and right shoulder keypoints, left and right hip keypoints, and left and right ankle keypoints, respectively, and takes the average of these two midpoints as the body's centroid. Simultaneously, the vector pointing from the midpoints of the hips to the midpoints of the shoulders represents the torso direction. Subsequently, the processor performs differential calculations on the centroid positions of adjacent frames to obtain the advancing speed along the gate's normal and the lateral contact speed along the gate's transverse direction, and calculates the degree of consistency between the torso direction and the gate's normal. This calculation reveals whether the object is continuously moving towards the gate in the direction of passage.

[0047] This formula corresponds to the calculation steps of propulsion speed, lateral contact speed and torso orientation consistency. The input is the human body centroid sequence and torso normal vector. It is used to solve the problem that the body facing and close contact state cannot be described by the target box trajectory alone. The output is multiple single-frame attitude dynamics indicators. The processor is the execution subject and the call time is after the coordinate mapping is completed.

[0048] ,

[0049] in, and These represent the coordinate components of the target's centroid in the normal and lateral directions of the gate, respectively, in meters; The time interval between adjacent frames, in seconds; This represents the trunk normal vector pointing from the midpoint of the hips to the midpoint of the shoulders; It is the normal basis vector in the local coordinate system; For the consistency of torso orientation; To prevent division by zero in the denominator, a preset smoothing term is used in this embodiment to avoid division by zero when the vector magnitude is close to zero. This parameter can be obtained through offline debugging according to numerical stability requirements, and its typical value is [value missing]. .

[0050] Subsequently, the processor extracts the forward probing amount and its differential rate of change of the ankle based on the normal lead of key points of the left and right ankles relative to the center of mass, which is used to identify short-term forward probing, retreat, and charging preparation movements. This step does not truncate negative values, thus preserving retreat movement information, which facilitates subsequent identification of reverse turning back.

[0051] In this embodiment, the single-frame feature dimension of the individual attitude feature matrix is ​​used to represent the core variables of the passage attitude retained in each frame within a time window, which can be determined by the effectiveness verification of risk identification; its typical value is 6, corresponding to propulsion speed, lateral contact speed, torso orientation consistency, ankle forward probing, ankle forward probing difference, and current center of mass normal distance. If The velocity term and the foot and ankle probing differential term are initialized to 0. When both shoulders or both hips are missing and the trunk orientation cannot be stably estimated, the valid value of the previous frame is used first, and it is set to 0 if two consecutive frames are missing. When both feet and ankles are missing, the foot and ankle probing differential uses the valid value of the previous frame, and it is set to 0 if two consecutive frames are missing.

[0052] In this embodiment, the normal propagation velocity sequence of target A in the most recent 5 frames is as follows: The lateral contact velocity sequence is The sequence of torso orientation consistency is as follows The forward probing sequence of the relative centroids of the left and right ankles is as follows: The corresponding foot and ankle probing differential sequence is: The processor stacks these quantities with the current centroid normal distance in chronological order to obtain... The calculation process of the above-mentioned multiple quantities constitutes the aforementioned individual attitude feature matrix.

[0053] Step S103: Construct a multi-person coupled tailing state

[0054] Based on the human centroid trajectory of each target in the individual posture feature matrix, the normal distance difference between adjacent targets is calculated; based on the lateral coordinates of the shoulder and hip key points of each target in the metric local coordinates, the lateral projection interval of each target is constructed, and the lateral projection overlap rate between adjacent targets is calculated according to the lateral projection interval; based on the human centroid trajectory, the propulsion direction vector of each target is calculated, and the propulsion direction similarity between adjacent targets is calculated according to the propulsion direction vector of each target; based on the normal distance difference, lateral projection overlap rate, and propulsion direction similarity, the trailing coupling strength is calculated in combination with a preset distance attenuation coefficient, and a preset number of neighboring targets with the highest coupling strength are selected for each target to form a frame-by-frame trailing coupling vector; the frame-by-frame trailing coupling vector is concatenated with the individual posture feature matrix frame by frame to obtain the human passage intention state tensor.

[0055] As a specific implementation method, the specific execution logic of the above process is as follows: For multiple targets in the same channel and the same processing cycle, the processor first determines the relative distance between front and rear targets based on the position difference of the centroid of each target in the gate normal direction; then, it constructs the lateral projection interval based on the extreme values ​​of the left and right shoulders and left and right hips in the lateral direction, and calculates the lateral projection overlap rate between the target pairs; at the same time, the processor obtains the propulsion direction vector based on the centroid difference of adjacent frames, and uses cosine similarity to measure whether the two targets maintain the same direction of propulsion.

[0056] To achieve tail-following coupling modeling, the preset distance attenuation coefficient in this embodiment is used to characterize the attenuation rate of tail-following association strength as the normal distance increases. This parameter can be obtained through offline fitting of historical tail-following events and normal parallel passage samples; in practical engineering applications, its typical value is 0.45, and it can usually be configured within the range of 0.30 to 0.80. After the similarity between the aforementioned lateral projection interval and the propulsion direction is determined, the processor calculates the tail-following coupling strength. This formula corresponds to multi-person tail-following relationship modeling, with the input being the normal distance difference, lateral projection overlap rate, propulsion direction similarity, and distance attenuation coefficient. It is used to solve the problem that traditional single-target features cannot identify multi-person coupling risks such as card borrowing and gate rushing. The output is the frame-by-frame tail-following coupling strength. The execution entity is the processor, and the timing of the call is after the extraction of single-target attitude features is completed in each processing cycle.

[0057] ,

[0058] in, Indicates the first Target in frame With the goal The tailing coupling strength is such that a larger value indicates that the two are more likely to form a close-range tailing relationship in the same direction. This represents the difference in normal distance between the centroids of two targets, expressed in meters. The preset distance attenuation coefficient; This is the horizontal projection overlap rate, with a value ranging from 0 to 1; The cosine similarity is used for the propulsion directions. When the propulsion directions of two targets are inconsistent, the processor determines the cosine similarity by taking... Suppress false triggering of following relationships by reverse or lateral motion. If the advancing speed of a target is lower than the static determination condition, the similarity is directly set to 0 to prevent static jitter from being misjudged as same-direction following.

[0059] The aforementioned preset static threshold is used to determine whether the target is in a valid moving state. It can be obtained by statistically analyzing the upper limit of velocity noise of stationary personnel samples under camera shake conditions. Its typical value is 0.05 m / s, and it can usually be adjusted between 0.02 m / s and 0.10 m / s in practical engineering applications. If the velocity modulus is lower than this value, the similarity of the target's propulsion direction is recorded as 0.

[0060] After calculating the trailing coupling strength, the processor selects the two nearest neighbors with the strongest coupling strength for each target in each frame, padding any missing parts with zeros, and concatenates them with the pose features of the same frame. If there is only a single target, the trailing coupling vector is set to zero; if the lateral projection intervals have no intersection, the overlap rate is set to zero; when the number of effective neighboring targets is less than two, the missing positions are padded with zeros. Thus, the processor obtains the human traffic intention state tensor that varies over time.

[0061] In this embodiment, the normal distance difference between target A and target B in the 5th frame is 0.22m, the lateral overlap rate is 0.83, and the propulsion direction cosine similarity is 0.94. Substituting these values ​​into the aforementioned relationships yields the trailing coupling strength. The coupling value of another nearby target is 0.114, therefore the trailing coupling vector in frame 5 is... The processor generates coupling vectors frame by frame for the most recent 5 frames and then... By splicing the corresponding frames, the final result is The process of concatenating the frame-by-frame trailing coupling vector with the individual pose feature matrix constitutes the aforementioned human traffic intention state tensor.

[0062] Step S2: Based on the human body passage intention state tensor and identity verification results, generate a visual intention risk score that integrates identity and behavior.

[0063] like Figure 3As shown, this step is used to map static identity results, short-term behavioral intentions, and structural anomaly levels into a risk space that can directly drive access control actions.

[0064] Step S201: Construct static identity constraint vector

[0065] The system receives identity verification results from card readers, visitor reservation platforms, and faculty databases, and encodes these results into a static identity constraint vector. This vector includes document matching status, passage time matching status, authorized channel matching status, and peer verification status. Based on the current target's identity verification time, the system searches for adjacent targets in the security zone in front of the same channel gate within preset time windows before and after the current target. The peer verification status is determined based on the trailing coupling strength, independent identity verification results, and similarity of the advancing direction of these adjacent targets. The static identity constraint vector is extended to the time dimension and then concatenated with the human body passage intention state tensor to obtain a fusion input tensor. This fusion input tensor is then used to generate the visual intention risk score.

[0066] In this embodiment, the processor receives the card swipe result from the card reader, the reservation time slot and reservation identity information from the visitor reservation platform, and the identity attribution information from the faculty or student database. It then encodes the document matching status, access time matching status, authorized channel matching status, and peer verification status into a four-dimensional static identity constraint vector. Here, the preset time windows before and after are used to determine whether there are any peers in the vicinity of the current target that form a synchronous proximity relationship with it. Their engineering meaning is a short-term proximity retrieval range centered on the current card swipe or QR code scan time, which can be determined by jointly calibrating the access control interaction delay and the time it takes for people to pass through the crowd. In this embodiment, a typical value is the number of preset time windows before and after. In practical engineering applications, it can typically be configured between 0.3s and 1.0s.

[0067] To identify abnormal peers, this step introduces the aforementioned peer verification state, which represents whether there is a peer object that has not completed independent identity verification but maintains a same direction, close proximity, and close contact with the current target. This state is typically determined based on the tailing coupling result, the independent identity verification result of adjacent targets, and the similarity of their advancing directions. To clarify the triggering strength of peer relationships, the preset peer coupling threshold in this embodiment is used to measure whether adjacent targets reach the coupling level that should trigger abnormal peer judgment during short-term close contact. This threshold can be obtained through offline testing with labeled tailing samples, with a typical value of 0.45, and can usually be configured between 0.35 and 0.60. If, within the preset time window before and after, another target has a coupling strength greater than 0.45 for at least two consecutive frames, and this target has not completed independent identity verification or the verification failed, and the similarity of the advancing directions of the two targets is greater than 0.6 in at least one frame, then the peer verification state is set to mismatch.

[0068] In this embodiment, the processor receives a message that target A has successfully swiped their card, but the card belongs to student C. Based on the facial recognition result, the document matching status is 0. Students are allowed to pass during the current time period, so the passage time matching status is 1. The current channel authorization status is 1. Within the 0.5s retrieval time window before and after the card swipe, target B satisfies a trailing coupling strength greater than 0.45 with A for three consecutive frames, and B has not independently completed identity verification; therefore, the peer verification status is 0. Thus, the static identity constraint vector of target A is... The processor copies it 5 times in the time dimension and... By concatenating the input tensors, we obtain the fused input tensor. The four components mentioned above, along with their temporal expansion process, constitute the aforementioned static identity constraint vector and fusion input tensor.

[0069] If the identity information is not read, the document matching status is recorded as 0; if the appointment information is empty, the passage time matching status and the authorized channel matching status are both recorded as 0; when a certain channel scenario does not require peer verification, the peer verification status is fixed at 1; when the identity interface is momentarily abnormal, not all identity components are directly interpreted as high risk, but only when there is continuous contact without independent verification is the peer verification status recorded as mismatch, thereby reducing the interference of a single network jitter on normal passage.

[0070] Step S202: Calculate the short window intent probability and structural anomaly score

[0071] The fused input tensor is input into the temporal feature extraction layer to extract the hidden state sequence corresponding to the continuous time window; a short window hidden state is generated based on the hidden state sequence; the short window hidden state is input into the intent mapping layer to output the intent probability vector, which corresponds to normal passage, probing and lingering, reverse turnaround, tailgating and gate rushing, and unauthorized entry.

[0072] The process of generating a visual intent risk score that integrates identity and behavior further includes: acquiring a pre-stored hidden state center of a normal passage sample and a preset scale parameter; calculating the distance between the short window hidden state and the hidden state center of the normal passage sample; and performing an exponential mapping based on the distance and the preset scale parameter to obtain a structural anomaly score.

[0073] In this embodiment, the processor integrates an input tensor input behavior intent prediction model. The model consists of a temporal feature extraction layer and an intent mapping layer, where the temporal feature extraction layer employs two layers of gated recurrent units. The first hidden dimension is 64, and the second hidden dimension is 32. Here, the engineering meaning of the two hidden dimensions is the feature capacity when performing temporal abstract encoding on the input short-window pose features and the fused identity information, respectively, which can be determined jointly by the training set validation loss and the upper limit of the deployment computing power; in this embodiment, their typical values ​​are 64 and 32, respectively. The batch size during the training phase is 128, the learning rate is 0.001, and the number of training rounds is 60. Training stops when the validation set loss decreases by less than 0.001 for five consecutive rounds. The training labels include five categories: normal passage, probing and lingering, reverse turning back, tailgating, and unauthorized entry, formed by manually labeled video recordings and synthetic samples of abnormal behavior.

[0074] In this embodiment, the hidden state center of the normal passage sample is used to characterize the central position of normal passage in the short window hidden space. It can be obtained offline by manually verifying normal passage segments without abnormalities from campus passage records for no less than 7 consecutive days. The engineering meaning of its update cycle is the frequency of offline re-estimation of the center vector and scale parameter, which can be determined based on the stability of the passage pedestrian flow distribution, with a typical value of 7 days. To ensure statistical stability, the threshold for the number of newly added normal samples is used to determine whether the recalibration conditions are met, and can be set by the minimum sample variance requirement, with a typical value of 500 segments.

[0075] In this embodiment, the preset scale parameter is used to characterize the dispersion of normal samples in the hidden state space. It can be obtained by multiplying the median squared Euclidean distance from the normal sample to the center by a scale correction factor. The scale correction factor, in engineering terms, represents the amount of correction to expand the tolerance range of the normal distribution. It is typically determined based on offline calibration using historical false alarm rates, and its typical value range is 1.5 to 2.5. The scale parameter calculated in this embodiment corresponds to... .

[0076] This formula corresponds to the calculation of structural anomaly scores. The inputs are the short window hidden state, the hidden state center of the normal passing sample, and the preset scale parameters. It is used to solve the problem that it is difficult to cover out-of-distribution abnormal samples when only using classification probabilities. The output is the structural anomaly score. The execution subject is the processor, and the timing of the call is after the behavior intent prediction model outputs the short window hidden state.

[0077] ,in, This is the structural anomaly score, with a value between 0 and 1. The hidden state vector of the current target is a short window with a dimension of 32; The hidden state center for normally passing samples; Preset scale parameters; This is a preset smoothing term. When the current sample moves away from the normal transit center, the exponential term decreases, thus increasing the structural anomaly score. This formula belongs to the anomaly mapping relationship established based on the normal sample distribution and is applicable when the hidden state center and scale parameters have been calibrated using offline data. When the value falls below a preset lower limit, the processor uses that lower limit to prevent abnormal score oversaturation. As an example, the preset lower limit can be between 0.10 and 0.20, and in this embodiment, it is preferably 0.12. If there is a recent change in the flow of people and three consecutive manual checks indicate an increase in false positives, the processor will be recalibrated in the next offline maintenance cycle. and .

[0078] In this embodiment, after the fused input tensor of target A is input into the model, the hidden state is output. The corresponding intention probability vector is These correspond to normal passage, tentative hesitation, reverse turnaround, tailgating, and unauthorized intrusion, respectively; simultaneously, based on the aforementioned structural anomaly score calculation relationship, the following is obtained: The aforementioned intention probability vector and structural anomaly score together constitute the behavioral intention prediction result described above.

[0079] When the input sequence length is less than a complete time window, i.e., only one frame of data, the processor skips the velocity difference term and only uses the current frame's pose, the relative position of the person and the gate, and the static identity constraint vector to output a simplified intent probability. After this processing, even if there is a short-term frame loss or the initial sequence is incomplete, the system can still maintain basic risk reasoning capabilities.

[0080] Step S203: Generate a visual intent risk score

[0081] The high-risk behavior items are obtained by extracting the tail-running intention probability, unauthorized intrusion intention probability, and reverse turn-back intention probability from the intention probability vector; the identity mismatch item is calculated based on the matching of each component of the static identity constraint vector; the high-risk behavior item, structural anomaly score, and identity mismatch item are multiplied by their respective preset fusion weights, and the multiplication results are summed to obtain the visual intention risk score.

[0082] In this embodiment, the processor performs a weighted fusion of high-risk behavior items, structural anomaly scores, and identity mismatch items. To balance the risk contributions from multiple sources such as tailgating, reverse turnaround, structural anomalies, and identity mismatch, this step introduces preset fusion weights. The engineering meaning of this parameter set is the relative influence ratio of different risk components in the final risk score. It can be obtained through offline verification set grid search, aiming to minimize the weighted sum of the high-risk event false negative rate and false interception rate. In this embodiment, the typical values ​​of the four preset fusion weights are 0.38, 0.18, 0.24, and 0.20, respectively, and their sum is 1. This weighted set is derived from offline verification of main gate data over seven consecutive days.

[0083] This formula corresponds to the fusion step of visual intent risk score. The inputs are the high-risk component in the intent probability vector, the structural anomaly score, and the static identity constraint vector. It is used to solve the problem that a single identity judgment or a single behavior classification is difficult to cover the complex risk. The output is a risk score that can be directly used for access control decision-making. The execution subject is the processor, and the timing of the call is after the intent prediction and anomaly score calculation are completed.

[0084] ,

[0085] in, For the goal Visual intention risk score; , and These represent the probabilities of the intent to tailgate, unauthorized entry, and reverse turnback, respectively. Structural anomaly score; The first in the static identity constraint vector One component; to These correspond to the preset fusion weights. This relationship reflects that when the probability of high-risk behavior increases, the degree of structural anomaly strengthens, or the mismatch of identity constraint components increases, the risk score rises synchronously. If the sum of the probabilities of the five types of intents is not 1 due to numerical truncation, the processor first performs renormalization; when all four identity components are missing, the identity mismatch item is not directly judged as the highest risk, but is assigned a neutral value of 0.5 to reduce the amplified impact of interface failure on the risk result.

[0086] In this embodiment, the preset release threshold is used to delineate low-risk areas that can be directly released. Its engineering meaning is the upper limit of risk that allows access control to perform release actions, which can be obtained through offline calibration of the release rate target from historical normal passage samples; its typical value is 0.35, and in practical engineering applications, it can usually be fine-tuned between 0.25 and 0.45. The preset interception threshold is used to delineate high-risk intervals that require immediate locking and alarm triggering, and can be obtained through offline calibration based on the first-round interception rate target of high-risk events; its typical value is 0.60, and it can usually be configured between 0.50 and 0.75. In this embodiment, target A is obtained by substituting it into the aforementioned formula. The value is greater than 0.35 and less than 0.60, therefore it enters the delayed review interval. The fusion process of the above-mentioned high-risk behavior items, structural anomaly scores, and identity mismatch items constitutes the aforementioned visual intent risk score.

[0087] Step S3: Based on the visual intent risk score, generate a spatially bound holographic warning label.

[0088] This step combines Figure 4 The purpose of this implementation is to map abstract risk outcomes into visual labels consistent with the on-site gate space, enabling on-duty personnel to directly locate risk objects and their handling locations within the mixed reality terminal.

[0089] Step S301: Calculate spatial anchor points and early warning levels

[0090] Based on the human centroid trajectory, the propulsion direction vector, and the current centroid normal distance, combined with the preset warning extrapolation distance lower limit, the preset warning extrapolation distance upper limit, and the preset label upward movement distance, the spatial anchor point of the holographic warning label is calculated; according to the visual intent risk score and the distance amplification relationship between the current centroid normal distance and the preset safety distance, a warning level index is generated in combination with a preset distance amplification coefficient, and the risk level is determined based on the warning level index; the spatial anchor point is mapped to the terminal display coordinate system through the transformation relationship from the gate local coordinate system to the mixed reality terminal display coordinate system.

[0091] In this embodiment, the processor obtains the propulsion direction based on the target's current position and the centroid difference of the two most recent frames, and sets the tag anchor point near a potential conflict location in front of the target. Simultaneously, the tag is moved upwards along the head direction to avoid obscuring critical parts of the human body. To this end, this step introduces a preset lower limit and a preset upper limit for the warning extrapolation distance. These represent the minimum and maximum forward movement distance of the tag in the object's propulsion direction, respectively, and can be obtained through on-site visualization experiments. In this embodiment, their typical values ​​are 0.10m and 0.25m, and are typically configurable within the range of 0.05m to 0.30m. The preset tag upward movement distance is used to ensure the tag is displayed above the human body without obscuring critical parts. This can be determined through mixed reality terminal display comfort experiments, with a typical value of 0.35m, and is typically set between 0.30m and 0.40m.

[0092] Meanwhile, the preset safety distance in this embodiment is used to characterize the close-range risk area that requires special attention in front of the turnstile. It can be determined based on the common gathering length in front of the turnstile and the visual reaction distance of security personnel. Its typical value is 1.20m, and in actual engineering applications, it can usually be reasonably configured between 0.8m and 1.5m. The preset distance amplification factor is used to enhance the sensitivity of the risk level index to distance when the target approaches the turnstile. It can be calibrated through on-site erroneous blocking and missed blocking balance tests. Its typical value is 0.60, and it can usually be configured between 0.30 and 0.80.

[0093] This formula corresponds to the calculation of the warning level index. The inputs are the visual intent risk score, the current centroid normal distance, the preset safety distance, and the preset distance amplification coefficient. It is used to solve the problem of different handling priorities for the same risk score at different distance positions. The output is the warning level index that takes into account both the magnitude of the risk and the degree of proximity. The execution subject is the processor, and the timing of the call is after the visual intent risk score is generated and before the label content is output.

[0094] ,

[0095] in, This is the early warning level index; Visual intent risk score; This is the preset distance magnification factor; The normal distance from the target centroid to the gate passage line, in meters; A preset safe distance, in meters. The closer the target gets to the gate, the more... The smaller the value, the larger the distance amplification term, indicating that objects closer to the gate require a higher priority for handling under the same risk score. If the target has already crossed the gate's passage line, then... Record it as 0; if the safety distance calibration is abnormal, the channel preset configuration will be invoked.

[0096] In this embodiment, the processor classifies risk levels according to the following level mapping rules: when When marked as low risk; The time is marked as medium risk; when The time stamp is marked as high risk. This risk level classification can be obtained through statistical analysis of offline main gate data over seven consecutive days.

[0097] In this embodiment, the normal distance from target A to the gate passage line is Substituting the aforementioned relationship, we get The target is mapped to high risk. Combining the extrapolation distance adaptive rule, the label for target A moves forward by approximately 0.20m and upward by 0.35m. Subsequently, the processor maps the spatial anchor point in local coordinates to the terminal's world coordinates through a transformation from the channel local coordinate system to the mixed reality terminal display coordinate system. This transformation is jointly determined by the ground calibration board and the terminal's spatial positioning results; when the terminal pose changes, only the terminal coordinate transformation is updated without altering the aforementioned risk calculation results. If the terminal pose update delay exceeds a preset time limit, the label position of the previous frame is temporarily frozen and marked as position held. As an example, the preset time limit can be between 0.10 s and 0.30 s, and in this embodiment, it is preferably 0.20 s. The calculation results of the aforementioned spatial anchor point and warning level index constitute the aforementioned spatial binding basis.

[0098] Step S302: Generate holographic warning tag content

[0099] Based on the dominant risk intent, static identity constraint vector, and current channel number in the intent probability vector, a holographic warning label containing object identity type, dominant risk intent, risk level, suggested action, and channel number is generated; the spatial anchor point is mapped to the terminal display coordinate system through the transformation relationship from the gate local coordinate system to the mixed reality terminal display coordinate system to obtain a spatially bound holographic warning label.

[0100] Furthermore, in conjunction with the application scenario of this invention, after obtaining the warning level index, static identity constraint vector, and current channel number, the processor will generate corresponding text content and label styles. For objects whose dominant risk intent is tailgating, the recommended action is "Pay attention to adjacent objects traveling in the same direction"; for objects whose dominant risk intent is unauthorized intrusion, the recommended action is "Keep the gate closed and manually verify"; for objects whose dominant risk intent is reversing, the recommended action is "Prompt them to leave the reverse channel". When two intents have the highest probability, the intent category with the higher security risk is prioritized as the dominant risk intent; when the identity type cannot be identified, the identity type in the label is fixed as "Unverified Object".

[0101] To ensure cross-frame display stability, the processor employs a nearest-neighbor trajectory consistency rule to maintain tag binding. Here, the upper limits of centroid distance, time interval, and velocity direction angle are used to determine whether the target in the current frame can be continuously bound to the tag in the previous frame. These limits can be determined through mixed reality display continuity tests and offline calibration of occluded scene recordings. In this embodiment, the typical value for the upper limit of centroid Euclidean distance is 0.40m, the typical value for the upper limit of time interval is 0.6s, and the typical value for the upper limit of velocity direction angle is... If the continuous occlusion time exceeds 0.6 seconds, the original label becomes invalid and the candidate generation process is restarted.

[0102] In this embodiment, the holographic warning label generated by the processor for target A is: "Unverified Object / Taking-After-the-Gate Risk / High Risk / Pay Attention to Adjacent Objects / Channel 1". The label color is red, and the font size is a fixed high-risk font size. The combination of the above-mentioned object identity type, dominant risk intent, risk level, suggested action, and channel number constitutes the aforementioned holographic warning label.

[0103] Step S4: Based on the visual intent risk score and the holographic warning label, execute the access control action management and mixed reality linkage output.

[0104] This step combines Figure 5 The execution process transforms the aforementioned risk identification results into synchronized outputs of access control actions and spatialized early warnings.

[0105] Step S401: Generate access control decision

[0106] The visual intent risk score is compared with a preset release threshold and a preset interception threshold, and a release control command, a delayed review command, or a lock alarm command is generated based on the identity verification result. When multiple targets exist in the same channel within the same processing cycle, the access control decision is generated based on the target that is closest to the gate passage line and has the highest visual intent risk score. The release control command, delayed review command, or lock alarm command is sent to the access control execution interface and an access control receipt is received. When any of the following conditions are met: access control receipt timeout; no receipt received continuously; interface abnormality; the gate closure priority degraded control state is maintained.

[0107] In this embodiment, the processor compares the risk score with a threshold: when When both the document matching status and the passage time matching status are 1, a passage control command is output; when When, output a delayed review instruction; when When the time is right, a lock alarm command is output. Here, the engineering meaning of the delayed review command is a control action that maintains the gate closed for a short period of time and waits for the next processing cycle to be reviewed. Its duration can be set jointly according to the camera processing cycle and the gate execution delay; in this embodiment, its typical value is 0.4s, and it can usually be configured in the range of 0.2s to 0.8s.

[0108] To ensure closed-loop access control execution, this step also introduces an access control acknowledgment timeout. This timeout is the maximum tolerable time for waiting for an acknowledgment after sending an access control command, calibrated based on the communication performance of the on-site access controller. In this embodiment, the typical upper limit for the access control communication and acknowledgment reception phase is 35ms. If a timeout occurs or no acknowledgment is received twice consecutively, the gate closure priority is maintained in a degraded control state. If the access control interface returns to a busy state, the control state of the previous cycle is maintained, and the command is retransmitted once in the next cycle. When network jitter causes the access control status to be unknown, open label content that contradicts "allowed" is not output.

[0109] In this embodiment, target A is calculated for the first time. Therefore, the processor outputs a delayed verification instruction to keep the gate closed. Awaiting review in the next processing cycle. In the next processing cycle, the normal distance between target A and target B decreases to 0.14m, and after recalculation, the result is... The threshold value exceeded the preset interception threshold of 0.60, therefore a lock alarm command was output. The access controller returned an execution confirmation within 34ms, and the system updated the current channel status to lock alarm. The generation process of the above-mentioned access control command, delayed verification command, lock alarm command, and degraded control status constitutes the aforementioned access control decision.

[0110] Without affecting the main process corresponding to the aforementioned claims, in one optional embodiment, the access control instruction frame includes an instruction number, a controlled gate number, a generation timestamp, a predicted effective timestamp, a control action type, a task security level, and an integrity check code. The access control execution terminal determines whether the control instruction is within the allowed execution window based on the generation timestamp and the predicted effective timestamp, and determines whether the control instruction has been truncated, tampered with, or out of order during transmission based on the integrity check code. After this processing, the access control action output has better execution consistency and anomaly tracking capabilities.

[0111] Step S402: Output the mixed reality overlay result and channel control actions

[0112] The spatial anchor point, risk level, tag content, and target bounding box coordinates of the holographic early warning tag are sent to the mixed reality terminal, and when the mixed reality terminal is not connected, they are sent to the monitoring screen to complete the synchronous output of access control actions and mixed reality early warnings.

[0113] In this embodiment, the processor sends the tag anchor point, color level, text content, and target bounding box coordinates to the mixed reality terminal worn by the personnel on duty via the mixed reality display interface, and simultaneously sends opening and closing control actions to the turnstile via the access control interface. If the mixed reality terminal is temporarily unconnected, the access control action is still retained, and the holographic warning tag is output to the duty screen. To ensure data chain consistency, the processor establishes a synchronization record in the order of "acquisition frame number - target number - channel number - tag number - access control instruction number" so that the tag status of the same object is updated after the access control receipt is returned. When both the mixed reality terminal and the access control interface are abnormal, the system at least retains the local duty screen display and the turnstile closing priority strategy.

[0114] In this embodiment, the processor outputs a red holographic warning tag to the mixed reality terminal and sends a lock alarm command to the access control controller, ultimately keeping channel 1 closed and triggering an audio-visual alert. After this coordinated output, the visual risk result no longer remains in the background scoring, but is transformed into an executable access control action for a specific channel and a spatialized warning result superimposed on the on-site view. The aforementioned synchronization process between the tag output and the access control action constitutes the aforementioned access control action management and mixed reality coordinated output.

[0115] In this embodiment, the timing test results for the entire link are as follows: the maximum time for camera acquisition and frame buffering is 40ms, the maximum time for pose extraction is 55ms, the maximum time for coordinate mapping and feature calculation is 18ms, the maximum time for intent model inference and anomaly score calculation is 22ms, the maximum time for risk fusion and tag generation is 10ms, the maximum time for access control communication and receipt reception is 35ms, and the maximum time for gate execution response is 80ms. The first five items are executed in a pipelined or parallel manner in most cycles, and the actual single-cycle risk decision-making time is controlled within 145ms, which can complete at least one review before the gate executes the opening action.

[0116] In a 4-hour continuous simulation of the digital twin main gate scenario, a total of 8620 passage events were generated, including 126 tailgating incidents and 84 unauthorized intrusion incidents. The control scheme was "card access control + manual video patrol". Under this simulation condition, the average advance warning time for tailgating incidents improved from 0.42 s to 1.31 s; the average advance warning time for unauthorized intrusion improved from 0.37 s to 1.18 s; and the channel false opening rate decreased from 3.4% to 1.1%. In a 30-minute semi-physical joint debugging of a single channel, 326 people passed through, including 12 cases of tailgating risk manually marked and 5 cases of reverse backtracking. The average latency from the processor to the access control locking command was 153 ms, and the average latency of the access control receipt was 31 ms. Of the 12 tailgating risks, 10 entered a delayed review or locking state before the gate opened. It can be seen that the aforementioned steps S1 to S4 form a complete closed loop from attitude modeling, risk fusion, spatial display to access control.

[0117] Example 2: Risk Assessment of Fusion in a Multi-Identity Visitor Access Scenario

[0118] This example describes the visitor-only passage at the west gate of the campus. Visitors are required to make an appointment in advance and enter the campus by verifying their information via QR code and identification. This scenario often involves complex behaviors such as people with appointments approaching the passage with unapproved individuals, attempting to enter after the designated time has passed, and returning to approach again after scanning the code. Relying solely on static identity verification can easily misclassify "approved visitors" as low-risk.

[0119] This embodiment uses the same overall process and core computational logic as Embodiment 1, the difference being that the model training data and channel parameters use independently calibrated results for the visitor channel. The sampling frame rate remains 25 fps, the processing cycle is 0.2s, and the time window length is 5 frames. The offline training samples include 12,000 normal visitor passage segments and 3,800 abnormal behavior synthesized segments. The behavior intent prediction model training batch size is 128, the learning rate is 0.001, the number of training epochs is 60, and training stops when the loss on the validation set decreases by less than 0.001 for 5 consecutive epochs. The normal passage center vector of the unsupervised anomaly detector is calculated from the mean of the hidden states of 12,000 normal samples, and the scale parameter... Take 0.42.

[0120] In this embodiment, the specific value of the aforementioned preset distance attenuation coefficient remains 0.45; the specific values ​​of the aforementioned preset fusion weights remain 0.38, 0.18, 0.24, and 0.20; the specific values ​​of the aforementioned preset release threshold and preset interception threshold are 0.35 and 0.60, respectively; the specific value of the aforementioned preset distance amplification coefficient is 0.60; the specific value of the aforementioned preset safety distance is 1.20m; and the specific value of the aforementioned preset same-path coupling threshold is 0.45. All of the above parameters were obtained through offline calibration using the visitor channel records for the past 7 days, and the configuration of the student main gate channel is not directly applied.

[0121] In this scenario, the identity verification result for target D is: the QR code is valid, the visitor's name matches, but the arrival time exceeds the appointment time window by 15 minutes, and there is an unappointed accompanying person E located 0.18m behind D. Therefore, the processor generates a static identity constraint vector. The reason for the mismatch in the peer verification status is: before and after scanning the code. Within the time window, D and E satisfy the trailing coupling strength greater than 0.45 for two consecutive frames, and E does not independently complete identity verification. After executing steps S101 to S103 as described in Embodiment 1, the processor obtains the human passage intention state tensor of target D. The tail coupling principal value between D and E is 0.547.

[0122] Subsequently, the processor continues to execute steps S202 and S203 as described in Embodiment 1. The initial calculation yields the intent probability vector of target D as follows: The structural anomaly score is 0.58, and the visual intent risk score is 0.536. Since this value exceeds 0.35 but does not reach 0.60, the system does not directly allow passage but maintains a delayed review state. In the next processing cycle, target D exhibits obvious reverse turning and forward probing actions, and its intent probability vector is updated to... The structural anomaly score was updated to 0.61, and the visual intent risk score was updated to 0.548. Although it did not exceed the interception threshold of 0.60, the dominant risk intent changed from tailgating to reversing. Therefore, the system continued to keep the gate closed and displayed a "Prompt to leave the reverse channel" message in the holographic warning label.

[0123] A 72-hour simulation was built in the visitor-only channel, generating 4360 visitor events, including 98 instances of tailgating using borrowed codes, 117 instances of overstaying, and 52 instances of reverse return. The control scheme was "static identity verification + infrared beam alarm". Under this simulation condition, the false negative rate of high-risk events decreased from 15.6% to 4.8%; the interception rate of tailgating using borrowed codes increased from 81.2% to 94.9%; and the average invalid interception rate decreased from 6.7% to 3.9%. Ablation comparison results on the same simulation data showed that after removing the multi-person tailgating coupling branch, the false negative rate of high-risk events was 8.9%; after removing the identity constraint branch, it was 10.7%; and after removing the structural anomaly branch, it was 6.5%. This indicates that when multi-person tailgating coupling, identity constraints, and structural anomalies are all included in the same risk score, the judgment of the composite risk of "legitimate identity but abnormal behavior" is more stable.

[0124] Subsequently, based on the above calculation and processing results, the processor continues to execute subsequent steps S3 to S4 as described in Embodiment 1 to complete operations such as generating holographic warning tags, outputting access control action linkage, and degrading abnormal interfaces. The specific execution logic is the same as in Embodiment 1, and will not be repeated here.

[0125] Example 3: Mixed Reality Holographic Early Warning and Linkage under Multi-Target Congestion Conditions During Evening Rush Hour

[0126] This embodiment focuses on a short-term congestion scenario in front of a single passage at the main gate during the evening rush hour after school, emphasizing the verification of the spatial binding display and on-site linkage effect. In this scenario, even if the background video algorithm has already issued an alarm, security personnel often find it difficult to quickly locate the target in a dense crowd. Traditional two-dimensional background pop-up alarms still require on-duty personnel to switch their attention back and forth between the monitoring screen and the actual scene.

[0127] This embodiment uses the same main process as Embodiment 1, but focuses on verifying steps S301, S302, and S402. The mixed reality terminal is a head-mounted display device with a field of view of [missing information]. It is connected to the processor via a local area network. There are 7 targets within 1.5m in front of the current channel, among which target F is on the right side of the queue and is exhibiting lateral insertion and forward probing movements.

[0128] In this embodiment, the aforementioned preset distance magnification factor is specifically set to 0.60; the aforementioned preset safety distance is specifically set to 1.20m; the aforementioned preset release threshold and preset interception threshold are specifically set to 0.35 and 0.60, respectively; the aforementioned preset warning extrapolation distance lower limit and preset warning extrapolation distance upper limit are specifically set to 0.10m and 0.25m, respectively; and the aforementioned preset tag upward movement distance is specifically set to 0.35m. These parameters were all obtained through on-site visualization experiments and offline calibration of single-channel evening peak hour tests.

[0129] After executing steps S101 to S103 as described in Embodiment 1, the processor obtains the temporal state of target F: the average normal propulsion velocity is 0.48 m / s, the average lateral contact velocity is 0.21 m / s, the torso orientation consistency in the last frame is 0.81, the tailing coupling principal value with the target on the left is 0.43, and the lateral overlap rate with the target in front is 0.79. Continuing to execute steps S201 to S203, the static identity constraint vector of target F is... The intention probability vector is The structural anomaly score was 0.46, and the visual intent risk score was 0.470. This result indicates that although target F has a legitimate identity, he / she exhibits a tendency to tailgate and attempt to breach the gate.

[0130] Subsequently, in step S301, the processor calculates that the normal distance of target F from the gate's passage line is 0.24m, and the warning level index is 0.611, mapping it to high risk. According to the adaptive extrapolation distance rule, the warning extrapolation distance of target F is approximately 0.22m; therefore, the processor sets the spatial anchor point approximately 0.22m in front of target F and 0.35m above its head. The transformation from the gate's local coordinate system to the mixed reality terminal coordinate system is initialized through the ground AprilTag calibration board and updated in real time by the terminal pose module when the staff turns their head. Then, in step S302, the processor generates the tag content: "Student / Taking the Gate Risk / High Risk / Pay Attention to Adjacent Passengers / Channel 2".

[0131] In step S402, the label is superimposed on the target F in the mixed reality terminal's field of view with a red border and a directional guide arrow. Simultaneously, the processor sends a delayed verification command to the gate, allowing the personnel on duty to locate the target and verbally issue a stop without leaving their line of sight. If the mixed reality terminal is temporarily offline at this time, the same content is simultaneously sent to the duty screen, and the delayed verification action of the access control system continues to be effective.

[0132] In a digital twin scenario during evening rush hour, with a pedestrian density of 68 people per minute and a simulation duration of 2 hours, the control scheme was "video analysis background pop-up alarms." Under these simulation conditions, the following results were obtained: the average target confirmation time for security personnel decreased from 2.14 seconds to 0.86 seconds; the false opening rate of the passage decreased from 2.9% to 0.8%; and the average time for initial verbal intervention for high-risk individuals was advanced by 0.93 seconds. In a small-sample on-site joint debugging exercise for 20 minutes during a single-passage evening rush hour, security personnel wearing mixed reality terminals handled 11 high-risk alerts, and in 9 of these cases, they were able to complete target confirmation within 1 second of the first tag appearing. In the same scenario, when only background pop-ups were used, the average time for target confirmation for the same 11 alerts was 2.01 seconds. This demonstrates that directly spatializing the risk results onto the on-site view can effectively shorten the cognitive chain from alarm generation to target location and personnel response.

[0133] Subsequently, based on the above calculation and processing results, the processor continues to execute the subsequent steps as described in Example 1 to complete operations such as access control receipt processing, continuous tag binding maintenance, and abnormal degradation control. Its specific execution logic is the same as that in Example 1, and will not be repeated here.

[0134] Example 4: A Mixed Reality School Access Control System Based on Visual Intent Prediction

[0135] When implemented as a system, the system includes a state tensor construction module, a risk score generation module, a holographic label generation module, and a linkage control output module.

[0136] The state tensor construction module is used to construct a gate-aligned human passage intention state tensor based on continuous video frames captured by a calibrated fixed-view camera. This module specifically includes a video access unit, a pose extraction unit, a gate coordinate mapping unit, and a multi-person coupling modeling unit. The video access unit receives video streams from each channel; the pose extraction unit extracts a sequence of two-dimensional skeletal key points; the gate coordinate mapping unit maps the key points to a gate-aligned normalized coordinate system and recovers the metric local coordinates; and the multi-person coupling modeling unit generates a trailing coupling vector and concatenates it with the individual pose feature matrix.

[0137] The risk score generation module generates a visual intent risk score that integrates identity and behavior based on the human access intent state tensor and identity verification results. Specifically, this module may include an identity encoding unit, a fusion input construction unit, a behavior intent prediction unit, an anomaly detection unit, and a risk fusion unit. The identity encoding unit generates a static identity constraint vector; the fusion input construction unit constructs a fusion input tensor; the behavior intent prediction unit outputs an intent probability vector; the anomaly detection unit outputs a structural anomaly score; and the risk fusion unit generates the visual intent risk score.

[0138] The holographic tag generation module generates spatially bound holographic warning tags based on the visual intent risk score. Specifically, this module may include a spatial anchor point calculation unit, a level mapping unit, a tag content generation unit, and a terminal coordinate mapping unit. The spatial anchor point calculation unit calculates the tag anchor point in local coordinates; the level mapping unit generates a warning level index and classifies risk levels; the tag content generation unit generates the object identity type, dominant risk intent, risk level, suggested action, and channel number; and the terminal coordinate mapping unit maps the spatial anchor point to the mixed reality terminal display coordinate system.

[0139] The linkage control output module is used to execute the access control action management and mixed reality linkage output of the aforementioned method based on the visual intent risk score and the holographic warning tag. This module specifically includes an access control decision unit, an access control receipt processing unit, and a mixed reality output unit. The access control decision unit is used to generate release control commands, delay review commands, or lock alarm commands; the access control receipt processing unit is used to maintain the downgraded control state with access control closure priority; and the mixed reality output unit is used to send the holographic warning tag to the mixed reality terminal or monitoring screen.

[0140] The data flow between the above modules is as follows: the state tensor construction module outputs a human access intent state tensor to the risk score generation module; the risk score generation module outputs a visual intent risk score to the holographic tag generation module and outputs the identity and behavior fusion results required for control decisions to the linkage control output module; the holographic tag generation module outputs spatial anchor points, risk levels, and tag content to the linkage control output module; the linkage control output module receives the access control receipt and writes back the execution status to maintain consistency between the control status and the tag status. This system architecture can support all processing flows in the aforementioned method embodiments.

[0141] The embodiments of the present invention have been described in detail above. Without departing from the technical principles of the present invention, those skilled in the art can make reasonable adjustments to the model structure, parameter configuration and execution logic according to specific application scenarios, data scale and business accuracy requirements. Such equivalent substitutions and adaptive adjustments all fall within the protection scope defined by the claims of the present invention.

Claims

1. A mixed reality-based campus access control method based on visual intent prediction, characterized in that, include: Step S1: Based on the continuous video frames captured by the calibrated fixed-view camera, construct the human passage intention state tensor aligned with the gate. Step S2: Based on the human body passage intention state tensor and identity verification results, generate a visual intention risk score that integrates identity and behavior; Step S3: Based on the visual intent risk score, generate a spatially bound holographic warning tag; Step S4: Based on the visual intent risk score and the holographic warning label, execute the access control action management and mixed reality linkage output.

2. The mixed reality access control method based on visual intent prediction according to claim 1, characterized in that, Constructing the human access intent state tensor aligned with the turnstile includes: Human pose extraction is performed on the continuous video frames to obtain the two-dimensional skeletal key point sequence of each target within a continuous time window, and the gate reference endpoint and gate physical width are read. Based on the gate reference endpoint, a gate alignment normalized coordinate system is established with the gate's horizontal and normal directions as the basis vectors. The two-dimensional skeleton key point sequence is mapped to the gate alignment normalized coordinate system and converted into metric local coordinates in combination with the gate's physical width. Based on the metric local coordinates, the trajectory of the human body's center of mass, propulsion speed, lateral contact speed, torso orientation consistency, foot and ankle forward probing, foot and ankle forward probing difference, and current center of mass normal distance are extracted according to the midpoints of the shoulders, the midpoints of the hips, and the key points of the feet and ankles to obtain the individual posture feature matrix.

3. The mixed reality access control method for schools based on visual intent prediction according to claim 2, characterized in that, The construction of the gate-aligned human passage intention state tensor also includes: Based on the human centroid trajectory of each target in the individual posture feature matrix, the normal distance difference between adjacent targets is calculated. Based on the lateral coordinates of the shoulder and hip key points of each target in the metric local coordinates, the lateral projection interval of each target is constructed, and the lateral projection overlap rate between adjacent targets is calculated according to the lateral projection interval. The propulsion direction vector of each target is calculated based on the trajectory of the human centroid, and the similarity of propulsion directions between adjacent targets is calculated based on the propulsion direction vector of each target. Based on the normal distance difference, lateral projection overlap rate and propulsion direction similarity, the trailing coupling strength is calculated in combination with the preset distance attenuation coefficient, and a preset number of neighboring targets with the highest coupling strength are selected for each target to form a frame-by-frame trailing coupling vector. The frame-by-frame trailing coupling vector and the individual pose feature matrix are concatenated frame by frame to obtain the human passage intention state tensor.

4. The mixed reality access control method for schools based on visual intent prediction according to claim 1, characterized in that, The generation of a visual intent risk score that integrates identity and behavior includes: Receive the identity verification results returned by the card reader, the visitor appointment platform and the faculty and staff database, and encode the identity verification results into a static identity constraint vector. The static identity constraint vector includes the document matching status, the passage time matching status, the authorized channel matching status and the peer verification status. Based on the current target's identity verification time, adjacent targets in the security zone in front of the same channel gate are retrieved within the preset time windows before and after, and the same-line verification status is determined according to the tail coupling strength, independent identity verification results and similarity of the advancing direction of the adjacent targets. The static identity constraint vector is extended to the time dimension and then concatenated with the human passage intention state tensor to obtain a fusion input tensor. The fusion input tensor is then used to generate the visual intention risk score.

5. The mixed reality access control method for schools based on visual intent prediction according to claim 4, characterized in that, The generation of the visual intent risk score that integrates identity and behavior also includes: The fused input tensor is input into the temporal feature extraction layer to extract the hidden state sequence corresponding to the continuous time window; Generate a short-window hidden state based on the hidden state sequence; The short window hiding state is input into the intent mapping layer, and the intent probability vector is output. The intent probability vector corresponds to normal passage, probing and lingering, reverse turnback, tailgating and gate rushing, and unauthorized intrusion.

6. The mixed reality access control method for schools based on visual intent prediction according to claim 5, characterized in that, The generation of the visual intent risk score that integrates identity and behavior also includes: Obtain the hidden state center and preset scale parameters of the pre-stored normal passage sample; Calculate the distance between the center of the short window hidden state and the center of the hidden state of the normal passage sample; An exponential mapping is performed based on the distance and the preset scale parameter to obtain the structural anomaly score.

7. The mixed reality access control method for schools based on visual intent prediction according to claim 6, characterized in that, The generation of a visual intent risk score that integrates identity and behavior specifically includes: The high-risk behavior items are obtained by extracting the tail-flying and gate-rushing intent probability, unauthorized intrusion intent probability, and reverse turn-around intent probability from the intent probability vector. Calculate the identity mismatch term based on the matching status of each component of the static identity constraint vector; The high-risk behavior item, structural anomaly score, and identity mismatch item are multiplied by their respective preset fusion weights, and the multiplication results are summed to obtain the visual intent risk score.

8. The mixed reality access control method for schools based on visual intent prediction according to claim 7, characterized in that, The generated spatially bound holographic warning tags include: Based on the human centroid trajectory, the propulsion direction vector, and the current centroid normal distance, combined with the preset warning extrapolation distance lower limit, the preset warning extrapolation distance upper limit, and the preset tag upward movement distance, the spatial anchor point of the holographic warning tag is calculated; Based on the visual intent risk score and the distance amplification relationship between the current centroid normal distance and the preset safety distance, a warning level index is generated in combination with the preset distance amplification coefficient, and the risk level is determined based on the warning level index. Based on the dominant risk intent, static identity constraint vector, and current channel number in the intent probability vector, a holographic early warning label containing object identity type, dominant risk intent, risk level, suggested action, and channel number is generated. The spatial anchor point is mapped to the terminal display coordinate system through the transformation relationship from the gate's local coordinate system to the mixed reality terminal display coordinate system, thus obtaining a spatially bound holographic warning label.

9. A mixed reality access control method for schools based on visual intent prediction according to claim 1, characterized in that, The output of the interaction between access control actions and mixed reality includes: The visual intent risk score is compared with a preset allow threshold and a preset block threshold, and an allow control command, a delayed review command, or a lock alarm command is generated in combination with the identity verification result. When multiple targets exist in the same channel within the same processing cycle, access control decisions are generated based on the target that is closest to the gate passage line and has the highest visual intent risk score. Send the access control command, delay verification command, or lock alarm command to the access control execution interface and receive access control receipts. If any of the following conditions are met: access control receipt timeout; no receipt is received continuously; interface is abnormal; then maintain the degraded control state with gate closure priority. The spatial anchor point, risk level, tag content, and target bounding box coordinates of the holographic early warning tag are sent to the mixed reality terminal, and when the mixed reality terminal is not connected, they are sent to the monitoring screen to complete the synchronous output of access control actions and mixed reality early warnings.

10. A mixed reality access control system for schools based on visual intent prediction, characterized in that, include: The state tensor construction module is used to construct a human passage intention state tensor aligned with the gate based on continuous video frames captured by a calibrated fixed-view camera. The risk score generation module is used to generate a visual intent risk score that integrates identity and behavior based on the human passage intent state tensor and identity verification results. The holographic tag generation module is used to generate spatially bound holographic warning tags based on the visual intent risk score; The linkage control output module is used to implement the steps of the mixed reality access control method based on visual intent prediction as described in any one of claims 1 to 9, according to the visual intent risk score and the holographic warning label.

Images