Barrier gate security monitoring system based on quantum laser radar and AI vision

Abstract

The invention relates to the technical field of security and protection monitoring, in particular to a barrier gate security and protection monitoring system based on quantum laser radar and AI vision. The method comprises the following steps: a quantum laser radar sensing unit performs three-dimensional distance measurement and imaging on vehicles in a barrier gate area by using a quantum laser radar, outputs initial target point cloud data, and performs space-time calibration on an initial point cloud data set in combination with quantum correlation characteristics and a scanning mechanism motion model to generate an anti-jitter point cloud data set; the AI visual perception unit collects a barrier scene image, identifies a vehicle bounding box by using a YOLOv8 model, and extracts a motion track of a vehicle in combination with a DeepSORT tracking algorithm; and the multi-modal data fusion unit fuses the target point cloud data subjected to space-time calibration with the recognition result based on a cross-modal attention mechanism. According to the method, the laser radar based on the quantum entanglement light source is adopted, and the quantum correlation noise base is effectively extracted through the two-photon counting matrix and quantum time / space correlation analysis.

Description

Technical Field

[0001] This invention relates to the field of security monitoring technology, and more specifically, to a gate security monitoring system based on quantum lidar and AI vision. Background Technology

[0002] Currently, barrier gate systems are widely used in parking lots, community entrances and exits, transportation hubs, and other scenarios, undertaking the important functions of vehicle traffic management and security control. Traditional barrier gate security systems mainly rely on visible light cameras combined with image recognition algorithms (such as license plate recognition and motion detection) for vehicle monitoring. However, under complex lighting and adverse weather conditions such as nighttime, strong light, rain, fog, and dust, the visual system is prone to problems such as image overexposure, decreased contrast, or deterioration of signal-to-noise ratio, resulting in a significant decrease in vehicle recognition rate, inaccurate motion trajectory tracking, and difficulty in reliably capturing dangerous behaviors such as driving against traffic, rushing through the gate, and abnormal parking, resulting in large security blind spots. To improve perception capabilities, conventional LiDAR is introduced to obtain three-dimensional spatial information. However, its mechanical scanning structure (such as rotating mirrors and galvanometers) is easily affected by vibration, resulting in point cloud jitter, spatial offset, and temporal drift. Moreover, the signal-to-noise ratio drops sharply when detecting low-reflectivity targets or at long distances, leading to false detections and missed detections, making it difficult to guarantee stable and reliable three-dimensional modeling. Furthermore, existing systems generally suffer from problems such as spatiotemporal asynchrony, feature asymmetry, and rigid fusion strategies in multimodal data fusion. The dense semantic information of visual images and the sparse point cloud data of LiDAR are difficult to align and complement efficiently, and there is a lack of adaptive mechanisms to dynamically adjust modal weights according to the scene, limiting the integrity and robustness of target representation. At the behavioral analysis level, traditional methods are mostly based on static rule judgments (such as "excessive timeout equals illegal parking"), lacking the ability to model the long-term dependencies of vehicle movement trajectories, and cannot accurately identify complex dynamic behaviors (such as tentative advancement, detours, and gate breach prediction). Moreover, the alarm response strategies are singular, making it difficult to achieve risk classification and differentiated handling, resulting in high false alarm rates or delayed responses. As security demands develop towards intelligence and refinement, the limitations of traditional systems in terms of perception dimensions, environmental adaptability, fusion efficiency, and intelligent decision-making are becoming increasingly prominent. Therefore, this paper designs a gate security monitoring system based on quantum LiDAR and AI vision. Summary of the Invention

[0003] The purpose of this invention is to provide a gate security monitoring system based on quantum lidar and AI vision, in order to solve the problems mentioned in the background art, such as weak perception capability in complex environments, low accuracy of multimodal data fusion, inaccurate identification of abnormal behavior, and insufficient intelligent response level of traditional gate security systems.

[0004] To achieve the above objectives, the present invention aims to provide a gate security monitoring system based on quantum lidar and AI vision, comprising: The quantum lidar sensing unit uses quantum lidar to perform three-dimensional ranging and imaging of vehicles within the gate area, outputs initial target point cloud data, and combines quantum correlation features and scanning mechanism motion model to perform spatiotemporal calibration on the initial point cloud dataset to generate an anti-shake point cloud dataset. The AI ​​visual perception unit acquires images of the barrier gate scene, uses the YOLOv8 model to identify vehicle bounding boxes, and combines the DeepSORT tracking algorithm to extract the vehicle's motion trajectory. A multimodal data fusion unit, which fuses spatiotemporally calibrated target point cloud data with recognition results based on a cross-modal attention mechanism; An abnormal event detection and behavior analysis unit analyzes the vehicle's travel trajectory and abnormal vehicle status based on the fusion results using a bidirectional long short-term memory network. The intelligent decision-making and barrier gate control unit executes the corresponding barrier gate control logic based on the analysis results.

[0005] As a further improvement to this technical solution, the quantum lidar sensing unit utilizes quantum lidar to perform three-dimensional ranging and imaging of vehicles within the barrier gate area, including the following steps: S1.1 Acquire the echo signal intensity sequence of the quantum lidar at a fixed sampling frequency, and generate a timestamp sequence using a time synchronization module. ; S1.2. Preprocess the acquired echo signal intensity sequence to extract valid echo points; based on the main peak time difference of the detected echo signal... Calculate the initial distance value ; S1.3. Combining the spatial calibration parameters of the lidar, the coordinates of each effective echo point are converted into three-dimensional spatial coordinates. An initial point cloud dataset is generated on the vehicle surface, and quantum correlation features are extracted from the echo signal. Combined with the motion model of the scanning mechanism, the initial point cloud dataset is spatiotemporally calibrated to generate an anti-shake point cloud dataset.

[0006] As a further improvement to this technical solution, in step S1.3, quantum correlation features are extracted from the echo signal, and combined with the motion model of the scanning mechanism, spatiotemporal calibration of the initial point cloud dataset is performed, including the following steps: S1.31. Using the correlated photon pair signal emitted by the quantum entangled light source, quantum correlation analysis is performed on the echo signal to extract the quantum correlation noise floor in the echo signal; S1.32. Establish a six-degree-of-freedom motion error model for the rotating mirror scanning structure in quantum lidar; S1.33. Based on the real-time sampling data from the IMU inside the lidar, calculate the amplitude of the additional angle disturbance caused by mechanical jitter for the error term that directly affects the scanning angle. and frequency And a time-varying perturbation model of the scanning angle is constructed; S1.34 Constructing an angle-time transfer function based on a time-varying perturbation model This is used to describe the dynamic relationship between mechanical jitter and the laser emission angle and sampling timing. S1.35. Construct a state vector based on the quantum correlation noise basis and the time-varying perturbation model, and use extended Kalman filtering to perform spatiotemporal joint estimation to generate the calibration result after spatiotemporal joint estimation. S1.36. Based on the calibration results, the initial point cloud dataset is reconstructed and its distribution is optimized to output the optimized anti-jitter point cloud dataset.

[0007] As a further improvement to this technical solution, the quantum correlation analysis of the echo signal involves the following specific steps: constructing a two-photon counting matrix using photon pair signals emitted by a quantum entangled light source. And calculate quantum time correlation Correlation with quantum space Quantum time correlation Correlation with quantum space As a characteristic of quantum correlation And based on quantum correlation characteristics Establish a noise separation criterion function to extract the quantum correlation noise basis. .

[0008] As a further improvement to this technical solution, the AI ​​visual perception unit includes a visual acquisition module and a visual perception module; The visual acquisition module acquires a continuous image sequence of the barrier gate area at a fixed frame rate, and performs time-series numbering and frame synchronization processing on the continuous image sequence to generate a structured image data stream. The visual perception module receives a structured image data stream, inputs each frame of the image data stream into the YOLOv8 model, and outputs detection results including vehicle category, bounding box coordinates, and confidence level; and uses the DeepSORT multi-object tracking algorithm to generate continuous motion trajectory information of the vehicle.

[0009] As a further improvement to this technical solution, the multimodal data fusion unit fuses target point cloud data with recognition results based on a cross-modal attention mechanism, including the following steps: S2.1 Represent the anti-jitter point cloud dataset as a point cloud feature sequence The detection results are represented as image feature sequences. and the point cloud feature sequence With image feature sequences Perform normalization and time synchronization processing; S2.2, Extrinsic parameter matrix based on lidar Map the point cloud coordinate system to the camera coordinate system; And through the camera intrinsic matrix Projecting onto the image plane yields two-dimensional pixel coordinates. ; S2.3. Based on the spatial mapping relationship from the point cloud coordinate system to the image coordinate system, establish a spatial correspondence between the projected points of the point cloud and the target detection boxes in the image, forming a preliminary cross-modal pairing set. ; S2.4, Set cross-modal pairings Point cloud feature sequences With image feature sequences Input a Transformer module based on a dynamic modality equalization attention mechanism, output fused features. .

[0010] As a further improvement to this technical solution, in S2.4, the cross-modal pairing set is... Point cloud feature sequences With image feature sequences Inputting a Transformer module based on a dynamic modality equalization attention mechanism includes the following steps: S2.41 Calculate point cloud feature sequences based on the number of point cloud and image features, spatial distribution uniformity, and area of ​​the region. With image feature sequences Effective information density; S2.42. Cross-modal fusion is performed through a two-level attention mechanism of intramodal self-attention and balanced cross-modal attention; S2.43, Point cloud feature sequences With image feature sequences Perform adaptive feature sampling and enhancement; S2.44. Combining the two-level attention mechanism and adaptive feature sampling results, the final fused feature is output. .

[0011] As a further improvement to this technical solution, in S2.42, cross-modal fusion is performed through a two-level attention mechanism of intra-modal self-attention and balanced cross-modal attention, involving the following specific steps: Point cloud feature sequences and image feature sequences Intramodal self-attention calculations are performed separately, and queries are generated. ,key Sum Then, softmax weighting is applied to obtain the enhanced single-modal features of the point cloud. and image single-modal features ; Through modal equalization gating Dynamically adjusted enhanced single-modal features of point clouds and image single-modal features The contribution weights are determined, and element-wise weighting is used to form a balanced cross-modal attention output.

[0012] As a further improvement to this technical solution, the abnormal event detection unit analyzes the vehicle's travel trajectory and abnormal vehicle status, including the following steps: S3.1, Fusing features from consecutive time points The vehicle travel trajectory feature sequence is formed by combining them in chronological order. ; S3.2, Sequence of vehicle travel trajectory features In a bidirectional long short-term memory network, the historical trend and future prediction of vehicle motion are captured through forward and backward memory units, respectively, and time-related feature vectors are extracted. ; S3.3, A lightweight classification sub-network is connected to the output layer of the bidirectional long short-term memory network, based on time-related feature vectors. Output the vehicle's state label at the current moment to obtain the complete vehicle state sequence. ; S3.4, Based on vehicle state sequence With trajectory feature sequence It identifies abnormal behavior patterns of vehicles in the spatiotemporal dimension and generates probability distributions of abnormal event types through a Softmax classification layer; S3.5. Risk levels are determined based on the probability distribution of abnormal event types.

[0013] As a further improvement to this technical solution, the intelligent decision-making and barrier gate control unit executes the corresponding barrier gate control logic based on the analysis results, involving the following specific steps: receiving the analysis results from the abnormal event detection unit, parsing and mapping them to the preset control strategy, and performing intelligent control of the barrier gate based on the preset control strategy.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention relates to a security monitoring system for a barrier gate based on quantum lidar and AI vision. It employs a lidar based on a quantum entangled light source, effectively extracting the quantum correlation noise floor through a two-photon counting matrix and quantum temporal / spatial correlation analysis. A time-varying jitter correction model is constructed by combining IMU data and a six-degree-of-freedom motion error model. Extended Kalman filtering (EKF) is used for spatiotemporal joint estimation to dynamically calibrate the point cloud offset caused by mechanical scanning jitter, generating a high-quality, jitter-resistant point cloud dataset. This method not only suppresses the false detection and drift problems of traditional lidar under vibration, lighting changes, or weak signal conditions, but also fully leverages the advantages of quantum sensing in signal-to-noise ratio and ranging accuracy. Therefore, it can stably acquire the three-dimensional structural information of vehicles even in complex environments such as rain, fog, night, and strong light, significantly improving the system's environmental perception robustness and spatial modeling accuracy in barrier gate scenarios.

[0015] 2. This invention relates to a security monitoring system for a barrier gate based on quantum lidar and AI vision. It introduces a dynamic modal equalization attention mechanism, comprehensively considering the effective information density of point cloud and image features during cross-modal fusion. Adaptive weighted fusion at the feature level is achieved through intra-modal self-attention enhancement and gated equalization of cross-modal attention. Combining appearance features and trajectory information provided by YOLOv8 and DeepSORT, and precise three-dimensional motion trajectories provided by quantum lidar, the system constructs a spatiotemporally consistent fusion feature sequence. Furthermore, bidirectional LSTM is used to model long-term vehicle traffic behavior, and a lightweight classification network and Softmax risk assessment module are used to accurately identify and classify abnormal events such as illegal parking, wrong-way driving, and gate breaching into five risk levels (low to extremely high risk). Finally, the intelligent decision-making unit executes differentiated control strategies (such as delayed gate lowering, alarm, and gate locking) based on the risk level, realizing intelligent and refined security management with a closed-loop system from "perception—fusion—understanding—decision," significantly improving the automation level and security response efficiency of the barrier gate system. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of the present invention; The meanings of the labels in the diagram are as follows: 1. Quantum LiDAR sensing unit; 2. AI visual sensing unit; 3. Multimodal data fusion unit; 4. Abnormal event detection and behavior analysis unit; 5. Intelligent decision-making and barrier gate control unit. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example: Please refer to Figure 1 As shown, a security monitoring system for gates based on quantum lidar and AI vision is provided, including: The quantum lidar sensing unit 1 uses quantum lidar to perform three-dimensional ranging and imaging of vehicles within the gate area, outputs initial target point cloud data, and combines quantum correlation features and scanning mechanism motion model to perform spatiotemporal calibration on the initial point cloud dataset to generate anti-shake point cloud dataset. In this embodiment, the quantum lidar is a lidar based on a quantum entanglement light source, and it adopts a rotating mirror or galvanometer scanning mechanism. In this embodiment, the quantum lidar sensing unit 1 uses quantum lidar to perform three-dimensional ranging and imaging of vehicles within the barrier gate area, including the following steps: S1.1 Acquire the echo signal intensity sequence of the quantum lidar at a fixed sampling frequency (the echo signal intensity sequence of the quantum lidar refers to the sequence data formed by the change in intensity of each photon signal reflected back from the target object (i.e., the vehicle) over time after the lidar emits a laser pulse; it reflects the echo intensity information of the interaction between the laser and the object at different points in time), and generate a timestamp sequence using the time synchronization module. (To ensure the temporal consistency and traceability of the data, each frame of collected data is marked with a frame header, sequence encoded, and its integrity verified to form a data frame that conforms to a standardized structure, providing a foundation for subsequent digital processing.) S1.2. Preprocess the acquired echo signal intensity sequence, including digital filtering, background noise suppression, and peak detection, to extract effective echo points; based on the detected echo signal main peak time difference (i.e., the time delay of the strongest peak (main peak) in the echo signal relative to the transmitted pulse). Calculate the initial distance value ( In the formula, (at the speed of light), and store the results as data records in the form of time-distance pairs; The preprocessing process is as follows: First, high-frequency noise and interference are removed by digital filtering (such as low-pass or band-pass filters) to smooth the signal curve; then, background noise suppression is performed by calculating the background level using targetless areas or historical baseline signals and subtracting it from the echo signal to enhance the effective signal; finally, a peak detection algorithm is used to identify local maxima in the signal and extract the corresponding time position and intensity value. These peak points are the effective echo points. S1.3. Combining the spatial calibration parameters of the lidar (including scanning angle, installation attitude, and detection matrix), the coordinates of each effective echo point are converted into three-dimensional spatial coordinates. An initial point cloud dataset is generated on the vehicle surface, and quantum correlation features are extracted from the echo signal. Combined with the motion model of the scanning mechanism, the initial point cloud dataset is spatiotemporally calibrated to generate an anti-jitter point cloud dataset, which is used to address the spatial offset and temporal drift error problems caused by the mechanical jitter of the scanning mechanism. The spatiotemporal calibration of the initial point cloud dataset primarily addresses the problem that in complex outdoor scenarios like gate barriers where continuous mechanical vibration exists, traditional LiDAR systems suffer from spatial coordinate shifts, increased noise, and blurred object contours due to mechanical jitter of their scanning mechanisms (such as rotating mirrors and galvanometers) and environmental electromagnetic interference. This severely reduces the accuracy and reliability of vehicle detection and contour recognition. This invention introduces "quantum correlation features," an intrinsic, physical observation benchmark. Instead of passively "filtering" or "guessing" what constitutes noise, it actively utilizes the inherent correlation characteristics of signals generated by quantum entangled light sources to directly distinguish the real signal from noise / jitter errors in the echo. This "genetic" distinguishing ability gives it a strong ability to suppress various types of random noise and mechanical jitter at the same frequency as the signal, achieving purification from the information source. This allows for the generation of more accurate and more interference-resistant anti-jitter point clouds without sacrificing point cloud density and real-time performance. The process of extracting quantum correlation features from echo signals and combining them with a scanning mechanism motion model to perform spatiotemporal calibration on the initial point cloud dataset includes the following steps: S1.31. Using the correlated photon pairs emitted by the quantum entangled light source, quantum correlation analysis is performed on the echo signal to extract the quantum correlation noise floor in the echo signal. In this case, the quantum lidar uses the quantum entangled light source to emit pairs of photons, some of which illuminate the target and generate an echo, while the other part of the photons are retained as a reference. By performing quantum correlation analysis on the received echo photons and the reference photons, the real target signal can be distinguished from noise or random background light. Furthermore, quantum correlation analysis of the echo signal involves the following specific steps: constructing a two-photon counting matrix using photon pairs emitted by a quantum entangled source. (Two-photon counting matrix) For the The photon detection channel and the first Two-photon count between photon detection channels, i.e., the number of times photon pairs are detected simultaneously within a certain time window, where, For the first A quantum entangled light source emits photon signals through its emission channels. For the first Each photon receives the photon signal from the detection channel. This is a statistical function used to count the number of times photon pairs occur simultaneously (reflecting the temporal correlation of photon pairs) and to calculate quantum temporal correlation. Correlation with quantum space (In the formula, Covariance operator, used for quantization and The degree of linear correlation; a larger value indicates a stronger spatial correlation. In spatial coordinates The corresponding photon signal or detection channel, measured in photons per second. In spatial coordinates The corresponding photon signal or detection channel (unit: photons / second); quantum time correlation Correlation with quantum space As a characteristic of quantum correlation And based on quantum correlation characteristics Establish a noise separation criterion function to extract the quantum correlation noise basis. ; Among them, quantum time correlation Used to quantize echo signals at different time delays Quantum spatial correlation characteristics are used to analyze potential system noise or mechanical jitter components in signals; quantum spatial correlation Used to describe the spatial statistical correlation between different detector pixels or echo points, providing spatial constraints for distinguishing target signals from noise, and through quantum correlation features The system reference noise (quantum correlation features acquired under targetless conditions) is compared, and the correlation components corresponding to mechanical jitter or system noise are distinguished based on the noise separation criterion function, thereby extracting the quantum correlation noise basis; The noise separation criterion function is: ; In the formula, For the first The quantum noise determination value corresponding to each point (or echo sample) is used to determine whether that point belongs to the quantum correlation noise basis. This is the time weighting coefficient, with a value range of [0.3, 0.7], determined through expert experience. The spatial weighting coefficient, with a value range of [0.3, 0.7], is determined through expert experience. The second-order quantum time correlation distribution is given under an ideal static state (or a calibrated reference scenario). This represents the spatial covariance distribution among different pixels of the detector under a known reference scenario (e.g., a planar reflector or in-space calibration). S1.32. For the rotating mirror or galvanometer scanning structure in quantum lidar, establish a six-degree-of-freedom (6-DoF) motion error model to quantify and describe the attitude drift and minute displacement errors generated by the scanning mechanism in actual operation, including angular deviation and positional deviation, thereby providing precise physical constraints for the spatiotemporal calibration of point cloud data. ; In the formula, For scanning agencies Angular error in the axial direction, For scanning agencies Angular error in the axial direction, For scanning agencies Angular error in the axial direction, This is the angular error term, which describes the overall rotational deviation of the scanning mechanism. The position error term describes the overall translational deviation of the scanning mechanism. It is estimated from the IMU's accelerometer data through quadratic integration and is used for subsequent translational correction of the point cloud coordinates. For scanning agencies Displacement error in the axial direction For scanning agencies Displacement error in the axial direction For scanning agencies Displacement error in the axial direction For transpose operation, For time; S1.33. Based on real-time sampling data from the IMU (including a three-axis gyroscope and accelerometer) inside the lidar, the error term that directly affects the scanning angle (i.e., ), calculate the amplitude of the additional angular disturbance caused by mechanical vibration. and frequency (The calculation process is as follows: First, the angular velocity signal of the scanning mechanism is acquired using the three-axis gyroscope of the IMU inside the lidar, and the angle change curve is obtained by integration. At the same time, the linear acceleration measured by the accelerometer is combined to estimate the small displacement. Then, the angle change signal is analyzed in the frequency domain (such as Fast Fourier Transform, FFT) to identify the amplitude of the main vibration components as the jitter amplitude.) (This indicates the magnitude of the angular deviation caused by vibration in the scanning mechanism.) The main peak of the frequency response is the jitter frequency. (Represents the vibration frequency that changes with time) and constructs a time-varying perturbation model for the scanning angle (this model is the angle error in the six-degree-of-freedom model). (Detailed expression in a specific direction): ; In the formula, For the scanning agency at a certain time Total angular error (including static deviation and dynamic jitter). This refers to the reference angle error or static deviation, originating from the slowly varying angle deviation in the six-degree-of-freedom model. and static factors such as installation errors, The jitter phase describes the initial offset or phase change of mechanical jitter over time; The item is dynamic jitter; S1.34 Constructing an angle-time transfer function based on a time-varying perturbation model This is used to describe the dynamic relationship between mechanical jitter and the laser emission angle and sampling timing. Among them, the angle-time transfer function is constructed. The specific process is as follows: Quantum time correlation... Correlation with quantum space Introduce a perturbation model and generate a corrected perturbation model. (In the formula, The quantum noise weighting coefficient for the time dimension, with a value range of [0.1, 0.4], is determined through expert experience and is used to adjust the contribution of time correlation in the correction. This is a time-dimension conversion factor, measured in radians (rad), determined through expert experience. This is the spatial dimension conversion coefficient, measured in rad·second² / photon², determined through expert experience. Quantum time correlation The expected value represents the value at different time delays. The average quantum correlation of the next echo signal is used to identify temporal noise components. (In the formula, For time delay The maximum value, For time delay (minimum value) The quantum noise weighting coefficient for the spatial dimension is used to adjust the contribution of spatial correlation in the correction. Its value ranges from [0.1, 0.4] and is determined through expert experience. For quantum space correlation The expected value, describing the spatial average statistical correlation between different probe pixels or echo points (used to identify spatial noise components), is used to perform statistical constraints and physical consistency correction on noise components, thereby suppressing non-physical jitter signals that do not match quantum correlation characteristics. (In the formula, The area of ​​the spatial region (e.g., the region corresponding to the field of view); and the ideal scanning model (In the formula, The angle-time transfer function is obtained by superimposing the angular position that the rotating mirror or galvanometer should reach according to the designed scanning trajectory under the condition of no mechanical error or vibration disturbance, onto the corrected disturbance model. This function is used to characterize the dynamic change law of the angle of the scanning mechanism in actual operation; S1.35. Construct a state vector based on the quantum correlation noise basis and the time-varying perturbation model, and use the extended Kalman filter (EKF) for spatiotemporal joint estimation to generate the calibration result after spatiotemporal joint estimation. Specifically, the state vector is first constructed from the quantum correlation noise basis and the time-varying perturbation model. This includes dynamic parameters such as the scanning mechanism's angular deviation, positional deviation, and noise components; then, based on the point cloud coordinates observed by the lidar and the corresponding timestamps, the observed values ​​are... A nonlinear observation model is established with the state vector; then, the extended Kalman filter (EKF) is used for spatiotemporal joint estimation, including state prediction, linearized observation update, Kalman gain calculation, and state and covariance update, so as to dynamically correct the real-time attitude and ranging accuracy of the lidar considering quantum noise and mechanical disturbances; finally, the spatiotemporal joint estimation result corrected by EKF is output, which is used to reconstruct and optimize the initial point cloud, and achieve anti-jitter and spatiotemporally consistent point cloud calibration. S1.36. Based on the calibration results, the initial point cloud dataset is reconstructed and its distribution optimized to output an optimized, jitter-resistant point cloud dataset. Specifically, the spatiotemporal joint calibration results obtained through extended Kalman filtering (EKF) are first applied to each point in the initial point cloud dataset, and the original three-dimensional coordinates are... The transformation is performed based on the calibrated angle deviation and position correction to eliminate spatial offsets introduced by mechanical jitter and quantum noise. Then, the calibrated point cloud is resampled and its density optimized, including interpolating low-density areas to generate virtual points, downsampling high-density areas to balance the distribution, and combining the point's reflection intensity. With timestamp The data is weighted and then output as an optimized anti-shake point cloud dataset, ensuring that the point cloud is spatially uniform, continuous and highly accurate, providing reliable input for subsequent multimodal fusion and target recognition.

[0019] AI visual perception unit 2 acquires images of the barrier gate scene, uses the YOLOv8 model to identify vehicle bounding boxes, and combines the DeepSORT tracking algorithm to extract the vehicle's motion trajectory. In this embodiment, the AI ​​visual perception unit 2 includes a visual acquisition module and a visual perception module; The visual acquisition module acquires a continuous image sequence of the barrier gate area at a fixed frame rate, and performs time-series numbering and frame synchronization processing on the continuous image sequence (specifically: each frame image acquired at a fixed frame rate is assigned a unique frame number and timestamp according to the sampling time order, and the acquisition time of multiple cameras is aligned by the system clock or global synchronization signal; subsequently, the inter-frame interval is corrected according to the timestamp difference, and duplicate frames or frame loss anomalies are eliminated to ensure that the images acquired by different channels or different cameras correspond to the scene state at the same moment in the time dimension, thereby forming a continuous image sequence with time consistency and synchronization accuracy), generating a structured image data stream; The visual perception module receives a structured image data stream, inputs each frame of the image data stream into the YOLOv8 model, and outputs detection results including vehicle category, bounding box coordinates, and confidence level. It then employs the DeepSORT multi-object tracking algorithm to achieve cross-frame vehicle identity matching and trajectory association based on detection box features and ReID embedding vectors, generating continuous vehicle motion trajectory information. Simultaneously, it performs license plate sub-region localization and cropping on the detected vehicle areas, extracting the license plate position, size, and corresponding frame sequence to provide input data for subsequent license plate recognition and identity verification. The YOLOv8 model employs an end-to-end target detection architecture, consisting of an input layer, a backbone network, a feature fusion neck, and a detection head. The input layer receives image data that has undergone normalization and enhancement processing. The backbone network, based on an improved CSPDarknet structure, comprises a Convolutional Normalized Activation (CBS) module, a Feature Fusion (C2f) module, and a Fast Spatial Pyramid Pooling (SPPF) module, used to extract multi-scale semantic features. The feature fusion neck uses an FPN-PAN structure to fuse high-level semantics with low-level detail features at multiple scales. The detection head uses a decoupled structure, outputting target confidence, class probability, and bounding box coordinates for each scale feature map, and obtaining the final detection result through non-maximum suppression, achieving high-precision recognition and localization of vehicles in the input image. The process of generating continuous motion trajectory information of the vehicle is as follows: using the position, size, and confidence score of the vehicle bounding boxes detected by YOLOv8 as input, the appearance features of each detection box are extracted and an embedding vector is generated through a pre-trained ReID network. The system first characterizes the vehicle's appearance. Then, it calculates the appearance similarity and motion prediction distance between the detected target and the existing trajectory between adjacent frames, and uses a Kalman filter to predict the target's speed and position. Next, it employs a Hungarian matching algorithm to achieve optimal target allocation in both appearance and motion feature spaces, determining cross-frame identity associations. When a detected target is continuously matched, the trajectory state is updated; if no match is found for a long period, the trajectory is terminated. Finally, it outputs a continuous vehicle motion trajectory containing the target ID, position, speed, and time series, achieving cross-frame identity preservation and motion behavior tracking for the vehicle.

[0020] Multimodal data fusion unit 3 fuses the spatiotemporally calibrated target point cloud data with the recognition results based on a cross-modal attention mechanism; In this embodiment, the multimodal data fusion unit 3 fuses the target point cloud data with the recognition result based on a cross-modal attention mechanism, including the following steps: S2.1 Represent the anti-jitter point cloud dataset as a point cloud feature sequence ( For the first The three-dimensional spatial coordinates of each echo point (or target point). The echo signal strength corresponding to this echo point. For timestamps, The number of valid echo points in the current frame or the current scanning period) represents the detection result (i.e. the detection result of the AI ​​visual perception unit (2), which includes at least vehicle category, bounding box coordinates, and confidence level) as an image feature sequence. ( For the first in the image The bounding box coordinates of each target represent its position and size on the image plane. For vehicle category labels, For confidence level, This is the ReID embedding vector extracted by DeepSORT. (the number of targets detected in the current frame), and the point cloud feature sequence. With image feature sequences Normalization and time synchronization processes are performed to unify the two modal data to the same time scale and coordinate reference system. S2.2, Extrinsic parameter matrix based on lidar (The LiDAR coordinate system maps the point cloud coordinate system to the camera coordinate system. The point cloud data was originally collected in the LiDAR's own coordinate system and cannot be directly mapped to the camera image; this is achieved through rotation in the extrinsic parameter matrix.) Peaceful relocation Transforming the point cloud to the camera coordinate system ensures that the projected position of the same spatial point on the image plane is consistent with the visual features captured by the camera, thereby achieving cross-modal registration and providing a unified spatial reference for subsequent point cloud and image feature fusion, target recognition, and trajectory analysis. In the formula, : No. The three-dimensional coordinates of a point in the camera coordinate system; And through the camera intrinsic matrix Projecting onto the image plane yields two-dimensional pixel coordinates. This is used to achieve pixel-level correspondence between point clouds and images, providing a foundation for cross-modal feature fusion; S2.3. Based on the spatial mapping relationship from the point cloud coordinate system to the image coordinate system, establish a spatial correspondence between the projected points of the point cloud and the target detection boxes in the image, forming a preliminary cross-modal pairing set. It is used to spatially align the 3D point cloud features of the LiDAR with the 2D target features in the camera image, so that each point cloud point can correspond to its corresponding image target detection box, thereby providing a one-to-one pairing input relationship for subsequent cross-modal feature fusion and realizing the joint representation of point cloud and image information. S2.4, Set cross-modal pairings Point cloud feature sequences With image feature sequences Input a Transformer module based on a dynamic modality equalization attention mechanism, output fused features. ; in, ; In the formula, This is a cross-modal attention weight matrix, representing the attention contribution of image features to point cloud features, used for weighted fusion. , It is a sequence of ReID (Representational Identity) features, including the category, bounding box, and appearance of the image target. Learnable projection / transformation matrices are used to map image features to the same feature space dimension as point cloud features. The spatial and reflection feature sequence of point cloud points; Among them, the cross-modal pairing set Point cloud feature sequences With image feature sequences Inputting a Transformer module based on a dynamic modality equalization attention mechanism includes the following steps: S2.41 Calculate point cloud feature sequences based on the number of point cloud and image features, spatial distribution uniformity, and area of ​​the region. With image feature sequences Effective information density is used to measure the information carrying capacity of point clouds and image features in the current frame: ; in The number of point cloud features. The number of image features, The percentage of the total area of ​​all detection boxes to the image. The total area of ​​the image. The variance of the point cloud distribution density within the detection box (reflecting spatial uniformity). The spatial response variance of image features. Point cloud feature sequence Effective information density, Image feature sequence Effective information density, This is a normalization factor (used for dimensional normalization, with units consistent with signal strength). S2.42. Cross-modal fusion is performed through a two-level attention mechanism of intramodal self-attention and balanced cross-modal attention; S2.43, Point cloud feature sequences With image feature sequences Adaptive feature sampling and enhancement are performed to address the issue of sparse or dense features and optimize cross-modal fusion performance. Specifically, for sparse point clouds, importance resampling is used to improve the representation of low-density regions, while feature interpolation is performed on low-density regions to generate virtual point cloud features, thus increasing the number of sampled point clouds. ( (The adjustable balance factor has an empirical value range of 0.5 to 2.0). For dense image features, the top-K most representative features are selected through the first-level intramodal attention weights to reduce redundancy, improve cross-modal fusion efficiency, and make the number of features match the point cloud. Furthermore, this two-level attention mechanism primarily addresses the fusion challenge caused by the mismatch in feature density, geometric structure, and information dimension between LiDAR point clouds and camera images—two heterogeneous data sets—in the context of a gate system. Specifically, point cloud data is sparse and spatially unevenly distributed (e.g., dense on the sides of a vehicle but sparse on the top), while image data is dense but lacks precise depth. Simple stitching or unidirectional alignment can lead to information overload or mismatches, especially when the vehicle partially obscures the view, is positioned at close range, or when severe weather causes a decline in the quality of data from a particular sensor. In such cases, the performance of traditional fusion methods deteriorates sharply, failing to provide a reliable and consistent joint feature representation for subsequent behavior analysis. This mechanism achieves intelligent fusion that prioritizes optimization before balancing. Intramodal self-attention first allows the point cloud and image to "organize their internal information," strengthening the key features of their respective modalities and suppressing redundancy. The subsequent balancing cross-modal attention acts like an "intelligent mixing console," dynamically adjusting the contribution weights of the two modalities in the final fused feature based on their real-time effective information density. This mechanism can adaptively respond to sensor performance fluctuations and scene changes: when the camera is blinded by strong light, the system automatically relies on more reliable point cloud features; when the point cloud is sparse due to heavy rain, it trusts visual features more. Thus, in complex real-world environments, it consistently outputs robust, complementary, and highly discriminative fused features, greatly improving the accuracy of subsequent vehicle tracking and abnormal behavior analysis. Cross-modal fusion is achieved through a two-level attention mechanism consisting of intra-modal self-attention and balanced cross-modal attention, involving the following specific steps: In the cross-modal fusion process, the point cloud feature sequence is first processed. and image feature sequences Intramodal self-attention calculations are performed separately, and queries are generated. ,key Sum Then, softmax weighting is applied to obtain the enhanced single-modal features of the point cloud. and image single-modal features This achieves information integration and feature enhancement within a modality, thereby improving the quality of single-modal features and providing stable input for subsequent cross-modal interactions. Specifically, it involves first processing the point cloud feature sequence... and image feature sequences Generate query vectors through linear transformations respectively Key vector Sum value vector (in For point cloud or image feature sequences, , , (As a learnable matrix), then calculate the attention weights. (In the formula, Let be the dimension of the key vector. The transpose operation is used to perform a weighted summation of the value vectors, thereby obtaining the enhanced single-modal features of the point cloud. and image single-modal features This process can integrate information within a modality, highlight key features, and suppress redundancy, providing stable and high-quality input for subsequent cross-modal interactions (this stage uses intramodal self-attention to enhance single-modal features). Through modal equalization gating Dynamically adjusted enhanced single-modal features of point clouds and image single-modal features The contribution weights are determined, and a balanced cross-modal attention output is formed using element-wise weighting. This addresses the mismatch in the number and distribution of features between the two modalities. (This stage involves balanced cross-modal attention using gated weights to weight and fuse point cloud and image features. Balanced cross-modal attention is a weighting mechanism introduced during cross-modal feature fusion to dynamically adjust the point cloud feature sequence.) With image feature sequences (This addresses the issues of differences in the number of features between the two modalities, uneven spatial distribution, or mismatched information density.) ; In the formula, To balance cross-modal attention output, This is element-wise multiplication; Among them, the gate value Depend on and via the Sigmoid function (i.e.) ) Calculations show that ,in The parameters are learnable, with empirical values ​​ranging from [0.5, 3.0], and are trained end-to-end via backpropagation. S2.44. Combining the two-level attention mechanism and adaptive feature sampling results, the final fused feature is output. : ; In the formula, Layer normalization is a process that standardizes the input features along the feature dimension to stabilize training and enhance the fusion effect.

[0021] The abnormal event detection and behavior analysis unit 4 analyzes the vehicle's travel trajectory and abnormal vehicle status based on the fusion results using a bidirectional long short-term memory network (Bi-LSTM). In this embodiment, the abnormal event detection unit 4 analyzes the vehicle's travel trajectory and abnormal vehicle status, including the following steps: S3.1, Fusing features from consecutive time points The vehicle travel trajectory feature sequence is formed by combining them in chronological order. ; S3.2, Sequence of vehicle travel trajectory features The input is fed into a Bidirectional Long Short-Term Memory (Bi-LSTM) network, where forward and backward memory units capture the historical trend and future prediction of vehicle motion, respectively, and extract time-related feature vectors. (To achieve long-term dependency modeling of dynamic characteristics of vehicle travel trajectories, where, For the forward long short-term memory network in time The hidden state vector encodes historical information from the beginning of the sequence to the current time step. For backward long short-term memory networks in time The hidden state vector encodes the future context information from the end of the sequence to the current time step. The Bidirectional Long Short-Term Memory (Bi-LSTM) network consists of an input layer, a bidirectional recurrent hidden layer, and an output layer: the input layer receives a time-series sequence of vehicle traffic trajectory features. The intermediate layer contains LSTM units in two directions—a forward LSTM from... arrive Capturing historical dependencies, backward LSTM from arrive To capture future information, each LSTM unit contains a forget gate, an input gate, and an output gate to control the memory and forgetting of information. The forward and backward hidden states are concatenated or summed at each time step to form a time-dependent feature vector. The output layer is usually a fully connected layer or a classification layer, which maps the hidden state to the target output, such as the state label or the predicted value, to achieve bidirectional modeling and dynamic feature extraction of sequence data. S3.3. A lightweight classification sub-network (composed of fully connected layers and activation layers) is connected to the output layer of the Bi-LSTM network, based on time-related feature vectors. Output the vehicle's current state label, including types such as normal passage, deceleration and waiting, abnormal stop, and wrong-way or boundary crossing, thereby obtaining a complete vehicle state sequence. ; The lightweight classification subnetwork typically consists of an input layer, a few intermediate layers, and an output layer: the input layer receives time-related feature vectors output by Bi-LSTM or other feature extraction modules. The intermediate layer consists of one or two fully connected layers (with ReLU or other lightweight activation functions) to perform nonlinear transformation and dimensionality reduction on the features, thereby enhancing the feature discrimination ability. The output layer is a fully connected layer with a Softmax activation function, which maps the intermediate features to the probability distribution of vehicle state categories, such as normal passage, deceleration and waiting, abnormal stagnation, or driving in the wrong direction, thereby achieving state classification at each time step in the sequence while maintaining a simple network structure and low computational cost. S3.4, Based on vehicle state sequence With trajectory feature sequence It identifies abnormal behavior patterns of vehicles in the spatiotemporal dimension, and generates probability distributions of abnormal event types through a Softmax classification layer, thereby achieving accurate classification of behaviors such as illegal parking, driving in the wrong direction, rushing through gates, and abnormal parking. Specifically, continuous state labels and corresponding trajectory features are input into a model (LSTM / Transformer with temporal dependence) for spatiotemporal behavior pattern analysis. By learning the behavior change patterns of vehicles over continuous time, abnormal behavior patterns (including at least continuous illegal parking, driving against traffic, and rushing through the gate) are identified. A Softmax classification layer is then connected to the output layer to map the identification results into the probability distribution of each abnormal event type, thereby realizing the quantification and classification of abnormal vehicle behavior. S3.5. Risk levels are classified based on the probability distribution of abnormal event types; Specifically: First, analyze the probability distribution of abnormal event types output by Softmax. Find the anomaly type corresponding to the highest probability (In the formula, (This represents the total number of abnormal event types), and then the base risk value corresponding to that abnormal event type in the preset risk table. And combined with the duration of the event Weighted by historical frequency Calculate the comprehensive risk score ( It is a time-weighted function that uses logarithmic growth. In the formula, and These are adjustable parameters used to control the scale and growth rate of the function. Based on experience, in this embodiment... , (This can be adjusted according to the actual scenario); Finally, Compare with a predefined threshold range, for example Low risk Medium risk. High risk, This indicates an extremely high risk level, thus completing the risk classification.

[0022] The intelligent decision-making and barrier gate control unit 5 executes the corresponding barrier gate control logic based on the analysis results; In this embodiment, the intelligent decision-making and gate control unit 5 executes the corresponding gate control logic based on the analysis results, involving the following specific steps: receiving the analysis results (including anomaly judgment results, status labels, and risk levels) from the abnormal event detection unit 4, parsing and mapping them to a preset control strategy, and intelligently controlling the gate based on the preset control strategy; specifically: low risk (normal passage) corresponds to the gate remaining open or controlled according to the normal opening and closing cycle; medium risk (deceleration and waiting or short-term abnormal stop) corresponds to extending the gate opening time or providing a brief alarm prompt; high risk (reverse passage, gate rushing, or abnormal stay) corresponds to immediately closing the gate and triggering an alarm signal; extremely high risk (serious violations or multiple abnormal events) corresponds to locking the gate, sending a remote alarm, and calling video monitoring records. The mapping table or rule engine stores the correspondence between risk levels and action strategies. After receiving the risk level, the mapping table is automatically queried, the corresponding control action command is parsed, and the corresponding operation is executed through the gate control module to achieve automated safety control.

[0023] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A security monitoring system for gate barriers based on quantum lidar and AI vision, characterized in that, include: Quantum lidar sensing unit (1) uses quantum lidar to perform three-dimensional ranging and imaging of vehicles in the gate area, outputs initial target point cloud data, and combines quantum correlation features and scanning mechanism motion model to perform spatiotemporal calibration on the initial point cloud dataset to generate anti-shaking point cloud dataset. AI visual perception unit (2), the AI ​​visual perception unit (2) collects the scene image of the gate, uses the YOLOv8 model to identify the vehicle bounding box, and combines the DeepSORT tracking algorithm to extract the vehicle's motion trajectory; Multimodal data fusion unit (3), which fuses the spatiotemporally calibrated target point cloud data with the recognition result based on a cross-modal attention mechanism; An abnormal event detection and behavior analysis unit (4) analyzes the vehicle's travel trajectory and abnormal vehicle status based on the fusion results using a bidirectional long short-term memory network. The intelligent decision-making and gate control unit (5) executes the corresponding gate control logic based on the analysis results.

2. The barrier gate security monitoring system based on quantum lidar and AI vision according to claim 1, characterized in that: The quantum lidar sensing unit (1) uses quantum lidar to perform three-dimensional ranging and imaging of vehicles within the barrier gate area, including the following steps: S1.1 Acquire the echo signal intensity sequence of the quantum lidar at a fixed sampling frequency, and generate a timestamp sequence using a time synchronization module. ; S1.

2. Preprocess the acquired echo signal intensity sequence to extract valid echo points; based on the main peak time difference of the detected echo signal... Calculate the initial distance value ; S1.

3. Combining the spatial calibration parameters of the lidar, the coordinates of each effective echo point are converted into three-dimensional spatial coordinates. An initial point cloud dataset is generated on the vehicle surface, and quantum correlation features are extracted from the echo signal. Combined with the motion model of the scanning mechanism, the initial point cloud dataset is spatiotemporally calibrated to generate an anti-shake point cloud dataset.

3. The barrier gate security monitoring system based on quantum lidar and AI vision according to claim 2, characterized in that: In step S1.3, quantum correlation features are extracted from the echo signal, and spatiotemporal calibration of the initial point cloud dataset is performed in conjunction with the motion model of the scanning mechanism, including the following steps: S1.

31. Using the correlated photon pair signal emitted by the quantum entangled light source, quantum correlation analysis is performed on the echo signal to extract the quantum correlation noise floor in the echo signal; S1.

32. Establish a six-degree-of-freedom motion error model for the rotating mirror scanning structure in quantum lidar; S1.

33. Based on the real-time sampling data from the IMU inside the lidar, calculate the amplitude of the additional angle disturbance caused by mechanical jitter for the error term that directly affects the scanning angle. and frequency And a time-varying perturbation model of the scanning angle is constructed; S1.34 Constructing an angle-time transfer function based on a time-varying perturbation model This is used to describe the dynamic relationship between mechanical jitter and the laser emission angle and sampling timing. S1.

35. Construct a state vector based on the quantum correlation noise basis and the time-varying perturbation model, and use extended Kalman filtering to perform spatiotemporal joint estimation to generate the calibration result after spatiotemporal joint estimation. S1.

36. Based on the calibration results, the initial point cloud dataset is reconstructed and its distribution is optimized to output the optimized anti-jitter point cloud dataset.

4. The barrier gate security monitoring system based on quantum lidar and AI vision according to claim 3, characterized in that: The quantum correlation analysis of the echo signal involves the following specific steps: constructing a two-photon counting matrix using photon pair signals emitted by a quantum entangled source. And calculate quantum time correlation Correlation with quantum space Quantum time correlation Correlation with quantum space As a characteristic of quantum correlation And based on quantum correlation characteristics Establish a noise separation criterion function to extract the quantum correlation noise basis. .

5. The barrier gate security monitoring system based on quantum lidar and AI vision according to claim 1, characterized in that: The AI ​​visual perception unit (2) includes a visual acquisition module and a visual perception module; The visual acquisition module acquires a continuous image sequence of the barrier gate area at a fixed frame rate, and performs time-series numbering and frame synchronization processing on the continuous image sequence to generate a structured image data stream. The visual perception module receives a structured image data stream, inputs each frame of the image data stream into the YOLOv8 model, and outputs detection results including vehicle category, bounding box coordinates, and confidence level; and uses the DeepSORT multi-object tracking algorithm to generate continuous motion trajectory information of the vehicle.

6. The barrier gate security monitoring system based on quantum lidar and AI vision according to claim 1, characterized in that: The multimodal data fusion unit (3) fuses the target point cloud data with the recognition results based on a cross-modal attention mechanism, including the following steps: S2.1 Represent the anti-jitter point cloud dataset as a point cloud feature sequence The detection results are represented as image feature sequences. and the point cloud feature sequence With image feature sequences Perform normalization and time synchronization processing; S2.2, Extrinsic parameter matrix based on lidar Map the point cloud coordinate system to the camera coordinate system; And through the camera intrinsic matrix Projecting onto the image plane yields two-dimensional pixel coordinates. ; S2.

3. Based on the spatial mapping relationship from the point cloud coordinate system to the image coordinate system, establish a spatial correspondence between the projected points of the point cloud and the target detection boxes in the image, forming a preliminary cross-modal pairing set. ; S2.4, Set cross-modal pairings Point cloud feature sequences With image feature sequences Input a Transformer module based on a dynamic modality equalization attention mechanism, output fused features. .

7. The barrier gate security monitoring system based on quantum lidar and AI vision according to claim 6, characterized in that: In S2.4, the cross-modal pairing set is... Point cloud feature sequences With image feature sequences Inputting a Transformer module based on a dynamic modality equalization attention mechanism includes the following steps: S2.41 Calculate point cloud feature sequences based on the number of point cloud and image features, spatial distribution uniformity, and area of ​​the region. With image feature sequences Effective information density; S2.

42. Cross-modal fusion is performed through a two-level attention mechanism of intramodal self-attention and balanced cross-modal attention; S2.43, Point cloud feature sequences With image feature sequences Perform adaptive feature sampling and enhancement; S2.

44. Combining the two-level attention mechanism and adaptive feature sampling results, the final fused feature is output. .

8. The barrier gate security monitoring system based on quantum lidar and AI vision according to claim 7, characterized in that: In S2.42, cross-modal fusion is performed through a two-level attention mechanism of intra-modal self-attention and balanced cross-modal attention, involving the following specific steps: Point cloud feature sequences and image feature sequences Intramodal self-attention calculations are performed separately, and queries are generated. ,key Sum Then, softmax weighting is applied to obtain the enhanced single-modal features of the point cloud. and image single-modal features ; Through modal equalization gating Dynamically adjusted enhanced single-modal features of point clouds and image single-modal features The contribution weights are determined, and element-wise weighting is used to form a balanced cross-modal attention output.

9. The barrier gate security monitoring system based on quantum lidar and AI vision according to claim 6, characterized in that: The abnormal event detection unit (4) analyzes the vehicle's travel trajectory and abnormal vehicle status, including the following steps: S3.1, Fusing features from consecutive time points The vehicle travel trajectory feature sequence is formed by combining them in chronological order. ; S3.2, Sequence of vehicle travel trajectory features In a bidirectional long short-term memory network, the historical trend and future prediction of vehicle motion are captured through forward and backward memory units, respectively, and time-related feature vectors are extracted. ; S3.3, A lightweight classification sub-network is connected to the output layer of the bidirectional long short-term memory network, based on time-related feature vectors. Output the vehicle's state label at the current moment to obtain the complete vehicle state sequence. ; S3.4, Based on vehicle state sequence With trajectory feature sequence It identifies abnormal behavior patterns of vehicles in the spatiotemporal dimension and generates probability distributions of abnormal event types through a Softmax classification layer; S3.

5. Risk levels are determined based on the probability distribution of abnormal event types.

10. The barrier gate security monitoring system based on quantum lidar and AI vision according to claim 1, characterized in that: The intelligent decision-making and gate control unit (5) executes the corresponding gate control logic based on the analysis results, involving the following specific steps: receiving the analysis results of the abnormal event detection unit (4), parsing and mapping them to the preset control strategy, and intelligently controlling the gate based on the preset control strategy.

Images