Method and system for detecting and identifying fence crossing and object throwing of pedestrians in airport based on edge calculation

Abstract

The invention belongs to the field of intelligent security and edge computing, and relates to an airport pedestrian fence crossing and parabolic object detection and recognition method and system based on edge computing, and the system collects a video stream in real time through a data preprocessing module, and locally executes background modeling, moving target detection and multi-target tracking in an edge computer; the space-time behavior analysis module extracts a human body skeleton point sequence, performs modeling analysis by using a space-time diagram convolutional network, and identifies a pedestrian fence crossing behavior; the trajectory analysis and recognition module performs trajectory fitting on a high-speed small target in the video, recognizes a parabolic behavior and predicts a drop point area; the fusion decision processing module fuses the recognition results of the crossing behavior and the parabolic behavior, performs conjoint analysis and triggers an alarm; and the early warning linkage execution module pushes the alarm information to an airport security system in real time and drives perimeter defense hardware to perform linkage. According to the invention, localized real-time detection, accurate identification and rapid early warning of fence crossing and object throwing behaviors of pedestrians at the periphery of the airport can be realized.

Description

Technical Field

[0001] This invention relates to the fields of smart security and edge computing, specifically to a method and system for detecting and identifying pedestrians climbing over fences and projectiles at airports based on edge computing. Background Technology

[0002] Airport perimeter security is a critical aspect of ensuring airport operational safety, but traditional detection methods have significant limitations in dealing with complex and high-risk behaviors. To address this, invention patent CN114913654A discloses an airport perimeter intrusion pre-alarm processing device based on edge computing. Its features include: a central control processing module and an image intelligent early warning module, an intrusion alarm module, an automatic tracking module, an AI deep learning module, and an execution module. The central control processing module interacts and communicates with these modules. The automatic tracking module is electrically connected to a high-speed dome camera. The execution module is electrically connected to LED lights, audible and visual alarms, and broadcasting equipment. The image intelligent early warning module is electrically connected to a vibration sensor, and the execution module is electrically connected to a network camera, effectively improving the accuracy of intrusion alarms and environmental interference resistance.

[0003] The above-mentioned technical solutions have made progress in reducing the false alarm rate, but the following technical problems still exist: they cannot accurately identify the continuous action sequence of pedestrians climbing over fences and the instantaneous motion trajectory of parabolic objects, and the reliance on cloud data processing results in high response latency and high bandwidth consumption, making it difficult to meet the needs of real-time identification of high-risk behaviors.

[0004] In view of this, it is very necessary to provide an airport pedestrian fence climbing and object throwing detection and recognition method and system based on edge computing to solve the above-mentioned defects in the prior art. Summary of the Invention

[0005] The purpose of this invention is to solve the problem that it is difficult to accurately identify the continuous action sequence of pedestrians climbing over fences and the instantaneous motion trajectory of projectiles, which makes it difficult to meet the needs of real-time identification of high-risk behaviors. In view of the technical defects of the above-mentioned existing technologies, this invention provides a method and system for detecting and identifying pedestrians climbing over fences and projectiles at airports based on edge computing, so as to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for detecting and recognizing pedestrians climbing over fences and projectiles at airports based on edge computing, comprising the following steps: Step S1: The data preprocessing step involves real-time acquisition of video streams from high-definition surveillance cameras around the airport, local background modeling and moving target detection by the edge computer, and running a multi-target tracking algorithm. Step S2: The step of spatiotemporal behavior analysis is to use the human key point detection network to extract the human skeleton point sequence and construct a spatiotemporal graph convolutional network to identify continuous behavior patterns of climbing and traversing. Step S3: The trajectory analysis and identification step involves the edge computer extracting high-speed small targets and calculating their trajectories based on a trajectory fitting algorithm, distinguishing high-risk projectiles and estimating their landing points. Step S4: The step of fusion decision processing, the edge computer performs joint analysis on the climbing behavior recognition result and the parabolic trajectory recognition result locally, triggers alarm and generates standardized alarm data; Step S5: The step of early warning linkage execution. The edge computer pushes the compressed alarm information to the airport security system in real time and links with the perimeter defense hardware.

[0007] Secondly, the present invention also provides an airport pedestrian fence-climbing and object-throwing detection and recognition system based on edge computing, comprising: The data preprocessing module is used to acquire video streams in real time from high-definition surveillance cameras around the airport and perform background modeling, moving target detection and multi-target tracking locally on the edge computer. The spatiotemporal behavior analysis module is used to extract human skeleton point sequences through a human keypoint detection network and to model and analyze them using a spatiotemporal graph convolutional network to identify pedestrian behavior of climbing over fences. The trajectory analysis and recognition module is used to fit the trajectory of high-speed small targets in the video, identify parabolic behavior, and predict their landing area. The fusion decision processing module is used to fuse the recognition results of climbing behavior and throwing behavior, perform joint analysis, and trigger alarms. The early warning linkage execution module is used to push compressed alarm information to the airport security system in real time and drive the perimeter defense hardware to perform linkage.

[0008] The modules work together to achieve localized real-time detection, accurate identification, and rapid early warning of pedestrians climbing over fences and throwing objects around the airport perimeter.

[0009] The beneficial effects of this invention are as follows: This invention realizes end-to-end localized intelligent processing from video acquisition and behavior analysis to early warning linkage, effectively solving the problems of insufficient detection accuracy, slow response speed and large bandwidth consumption in the traditional cloud processing mode, and significantly improving the overall effectiveness of airport perimeter security.

[0010] This invention effectively avoids long-distance transmission of video data by acquiring video streams, modeling the background, and detecting moving targets, thus solving the latency problem caused by cloud transmission at the source and ensuring the real-time performance of subsequent processing.

[0011] This invention utilizes a human keypoint detection network to extract skeleton point sequences and combines it with a spatiotemporal graph convolutional network to model continuous action sequences, accurately identifying behavioral patterns and solving the technical problem that traditional methods cannot identify continuous action sequences.

[0012] This invention achieves accurate capture of the instantaneous trajectory of a projectile and prediction of its landing point by extracting the projectile and calculating its trajectory fitting, overcoming the technical deficiency of existing technologies that make it difficult to accurately identify projectile behavior.

[0013] This invention achieves intelligent fusion and concurrent detection of multimodal data by jointly analyzing the recognition results of vaulting behavior and parabolic trajectory, effectively reducing the false alarm rate of the system and improving the accuracy and reliability of high-risk behavior recognition.

[0014] This invention constructs a complete localized early warning and response closed loop by pushing compressed alarm information to the airport security system in real time and directly linking it with perimeter defense hardware, which significantly improves the system's real-time response capability and handling efficiency to high-risk behaviors.

[0015] Therefore, it is evident that the present invention has outstanding substantive features and significant progress compared with the prior art, and the beneficial effects of its implementation are also obvious. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 This is a flowchart of a method for detecting and recognizing pedestrians climbing over fences and projectiles at airports based on edge computing; Figure 2 This is a schematic diagram of an airport pedestrian fence-climbing and object-throwing detection and recognition system based on edge computing. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following implementation methods.

[0019] Example 1: like Figure 1 As shown in the figure, this embodiment provides a method for detecting and recognizing pedestrians climbing over fences and projectiles at airports based on edge computing, which includes the following steps: Step S1: The data preprocessing step involves real-time acquisition of video streams from high-definition surveillance cameras around the airport, local background modeling and moving target detection by the edge computer, and running a multi-target tracking algorithm. Step S2: The step of spatiotemporal behavior analysis is to use the human key point detection network to extract the human skeleton point sequence and construct a spatiotemporal graph convolutional network to identify continuous behavior patterns of climbing and traversing. Step S3: The trajectory analysis and identification step involves the edge computer extracting high-speed small targets and calculating their trajectories based on a trajectory fitting algorithm, distinguishing high-risk projectiles and estimating their landing points. Step S4: The step of fusion decision processing, the edge computer performs joint analysis on the climbing behavior recognition result and the parabolic trajectory recognition result locally, triggers alarm and generates standardized alarm data; Step S5: The step of early warning linkage execution. The edge computer pushes the compressed alarm information to the airport security system in real time and links with the perimeter defense hardware.

[0020] In step S1: Video streams are acquired in real-time from high-definition surveillance cameras at the airport perimeter and transmitted to an edge computer at the airport perimeter to avoid long-distance transmission delays caused by video transmission back to the cloud. The edge computer uses ViBe or KNN background modeling algorithms to establish a stable dynamic background model, removing background noise such as lighting changes and swaying leaves from the video to reduce invalid data uploads. Based on the foreground segmentation results, motion regions are extracted, and moving targets are detected using object detection algorithms. Simultaneously, the edge computer runs the DeepSORT multi-target tracking algorithm to continuously track the moving targets, transforming the input video into a stable, continuous sequence of target trajectories, ensuring real-time tracking and providing structured data support for subsequent attitude analysis and parabolic trajectory detection. After the above processing, the original video stream is transformed into a continuous image sequence containing foreground segmentation results, target detection boxes, and stable target trajectory annotations.

[0021] By processing video streams locally on edge computers, delays from long-distance transmission are avoided, providing structured data support for subsequent behavior analysis and trajectory detection, effectively reducing bandwidth consumption, and minimizing false detection areas.

[0022] The background modeling algorithm described uses either ViBe or KNN. ViBe (Visual Background Extractor) is a fast, lightweight algorithm suitable for edge computing scenarios, used for both background modeling and foreground detection. Its core idea is to maintain a randomly sampled background model set for each pixel and classify foreground / background based on the similarity between the pixel and samples in this set. During initialization, the algorithm randomly selects neighboring pixels from the first or previous frames to construct a background sample set for each pixel. In the detection phase, the difference between the current pixel value and several samples in the background sample set is calculated. If the number of matches is insufficient, the pixel is classified as moving foreground; otherwise, it is classified as background. To adapt to scene changes, ViBe uses a random update strategy, writing the value of the current pixel or neighboring pixels into the background model with a certain probability, allowing the model to gradually absorb the changed background information. ViBe's advantages include low computational cost, low memory usage, and a stable update mechanism, making it suitable for rapid detection of moving targets in real-time scenarios such as airport perimeters.

[0023] KNN (K-Nearest Neighbor) is a foreground detection algorithm based on nonparametric density estimation. It determines whether a pixel belongs to the background by performing a K-nearest neighbor search on historical pixel samples. The algorithm typically involves maintaining a sample set containing brightness and color values ​​from several historical frames for each pixel. In the current frame, the Euclidean distance between the pixel and historical samples is calculated to find the K nearest neighbors. If the difference between most of these neighbors and the current pixel is less than a set threshold, it is considered background; otherwise, it is considered foreground. KNN can adaptively describe complex backgrounds such as lighting changes, shadows, water surfaces, and swaying branches based on sample distribution. Its advantages include strong robustness, independence from specific background model shapes, and stable performance on non-Gaussian backgrounds, but it has a relatively high computational cost. Because OpenCV provides an optimized implementation, KNN is suitable for modeling and real-time detection of large-scale, gradually changing backgrounds in high-resolution airport surveillance videos.

[0024] The multi-target tracking algorithm uses the DeepSORT algorithm, which (Deep Learning-based SORT) is an improvement on the traditional real-time multi-target tracking algorithm SORT. It primarily enhances target re-identification capabilities by introducing deep learning technology, thereby improving tracking accuracy in complex environments. Unlike SORT, which uses Kalman filters and the Hungarian algorithm for target motion prediction and data association, DeepSORT adds a target appearance feature extraction module. It extracts high-dimensional feature vectors from the target's appearance using Inception or ResNet deep neural networks, enabling the algorithm to accurately track targets even when they are occluded, briefly disappear, or undergo rapid re-identification. Specifically, the DeepSORT workflow includes: identifying targets in an image and generating bounding boxes using a YOLO or Faster R-CNN target detection model; then, predicting the target's motion using Kalman filtering and combining historical motion trajectories to provide a preliminary estimate of the target's position in the next frame; next, extracting the target's appearance features using deep learning for target re-identification in subsequent frames; and finally, matching the predicted position with the actual detected target using the Hungarian algorithm to achieve accurate multi-target tracking. DeepSORT enhances the tracking capabilities of targets in complex dynamic scenes by integrating motion and appearance features. It also significantly improves the algorithm's robustness and accuracy, making it particularly suitable for complex scenarios such as dense crowds, changing lighting, and partial occlusion. In security applications such as airport perimeter security, DeepSORT can achieve accurate multi-target tracking, effectively reducing target identification loss and supporting timely detection of dangerous behaviors such as climbing over fences or throwing objects.

[0025] In step S2: An edge computer is used to perform spatiotemporal behavior analysis on the continuous image sequence obtained in step S1. A human keypoint detection network is invoked to extract key skeleton points from multiple parts of the human body, forming a frame-level skeleton structure, which is then connected into a temporal sequence of the human skeleton. The ST-GCN spatiotemporal graph convolutional network is used to model the action sequence, constructing a keypoint-based spatiotemporal graph model. Joint convolutions are performed on human actions in both spatial topology and temporal dynamics dimensions to achieve the analysis of actions approaching a fence. climb Accurate identification of this continuous vaulting behavior pattern. Relying on ST-GCN's ability to model long temporal actions, it can reduce the probability of misidentifying a single frame action as vaulting and supports real-time inference, outputting vaulting behavior recognition results, specifically accurate vaulting behavior sequences and related attributes including timestamps, target locations, and behavior categories.

[0026] Using high-definition infrared surveillance cameras operating at 30fps along the airport perimeter and an edge computer within a 50-meter range of these cameras (equipped with an NVIDIA Jetson AGX Orin GPU acceleration module), the system detects pedestrians climbing over the fence at a distance of 50 meters. The edge computer runs an HRNet network and an ST-GCN spatiotemporal graph convolutional network to infer the pedestrian's climbing behavior, achieving pedestrian climbing action recognition with an accuracy of 93.6%, a false alarm rate of less than 5%, and a single inference latency of less than 150ms. This improves the accuracy and real-time performance of perimeter behavior recognition. Spatiotemporal behavior analysis accurately identifies continuous climbing behaviors, capturing the temporal features of the actions to achieve high-precision, high-real-time climbing detection. The output results provide direct evidence for edge-end alarms and security decisions, significantly improving the accuracy and reliability of perimeter behavior recognition.

[0027] The human keypoint detection network uses either the OpenPose network or the HRNet network. The OpenPose algorithm is a real-time human pose estimation algorithm based on convolutional neural networks, specifically designed to detect keypoints of the human body, such as shoulders, elbows, and knees. This algorithm uses Part Affinity Fields technology to locate keypoints in the human body in an image and forms a skeleton structure by connecting these keypoints. OpenPose can detect not only the pose of a single human body but also the poses of multiple bodies simultaneously, exhibiting excellent accuracy and real-time performance. Its applications include security monitoring and motion analysis. OpenPose's advantage lies in its ability to efficiently extract human skeleton points from videos, generating frame-level skeleton structures, providing foundational data for subsequent tasks such as motion recognition and pose analysis.

[0028] HRNet (High-Resolution Network) is another deep learning model for human pose estimation. By fusing multi-resolution feature maps, it retains higher-resolution information, thereby improving the accuracy of keypoint detection. HRNet achieves continuous preservation and fusion of high-resolution features through a cross-layer, multi-scale network architecture, thus enhancing the accuracy and robustness of human pose estimation. HRNet is used to extract key skeleton points from videos and generate high-precision skeleton structures, providing strong support for subsequent action recognition and spatiotemporal behavior analysis.

[0029] The ST-GCN (Spatio-Temporal Graph Convolutional Network) is a graph convolutional neural network used for spatiotemporal data modeling, particularly suitable for processing spatiotemporally structured data such as human skeleton data. ST-GCN treats the human skeleton in each frame of video as a graph, where nodes represent individual skeletal points and edges represent connections between joints. Through convolution operations on the spatiotemporal graph, ST-GCN can capture the spatial relationships between skeletal points and dynamic behavioral patterns that change over time, thus enabling accurate recognition of complex actions. ST-GCN is used to model the extracted temporal data of the skeleton to recognize continuous complex behaviors such as climbing over fences and throwing objects, allowing action recognition to consider not only the current posture but also the continuity and temporal sequence of the action.

[0030] In step S3: the trajectory of the projectile is analyzed to obtain the trajectory recognition result. The specific operation is as follows: The YOLOv8 small target detection model is used to extract small targets from the continuous image sequence obtained in step S1, identifying high-speed projectiles, including bottles, stones, and drones. Kalman filtering is applied to the detected small targets to predict their next position, ensuring consistency in target identity across consecutive frames. Subsequently, the time-series position points of the small targets are input into a parabolic motion model, and Kalman filtering is used for trajectory fitting. The velocity, acceleration, and direction of the object's flight path are estimated, distinguishing high-risk projectiles from normal flying birds, falling leaves, and other naturally disturbed targets, and predicting their future landing positions. The output is a projectile trajectory recognition result including the target's motion trajectory, velocity, acceleration, and landing position.

[0031] Using high-definition infrared surveillance cameras at 30fps along the airport perimeter and an edge computer within 50 meters of the cameras (equipped with an NVIDIA Jetson AGX Orin GPU acceleration module), the edge computer detected a thrown object at the fence 50 meters away. It then ran a YOLOv8 small target detection model, a parabolic motion model, and Kalman filtering to infer the trajectory of the projectile, achieving projectile detection. The edge computer achieved a 91.2% accuracy rate in recognizing the trajectories of targets such as bottles, stones, and drones locally, with a trajectory landing point prediction error of less than 0.5 meters and a trajectory analysis latency of less than 180ms.

[0032] Trajectory analysis enables high-precision detection and trajectory fitting of small targets at the edge, accurately predicting the flight path and landing point of projectiles, providing a reliable basis for timely early warning of potential threats, improving the accuracy and response speed of dangerous object identification, and reducing false alarms. It effectively reduces false alarms and provides early warning of dangerous objects that may fall into the runway or sensitive areas.

[0033] The YOLOv8 small target detection model described above can better capture detailed information, thereby improving the detection accuracy of distant or small objects. This model is suitable for real-time monitoring systems, capable of quickly processing video stream data and outputting detection results. In airport perimeter monitoring, YOLOv8 is used to detect high-speed flying small targets from real-time video, providing crucial data support for subsequent trajectory analysis and threat identification.

[0034] The parabolic motion model is a physical model describing the trajectory of an object under the influence of gravity. Parabolic motion is a typical quadratic curve motion, in which the trajectory of an object, under ideal conditions without air resistance, is influenced by the initial velocity, the throwing angle, and gravitational acceleration. The object moves independently in two directions: horizontally, it moves at a uniform linear velocity, while vertically, it undergoes acceleration due to gravity. This model is commonly used to simulate the trajectory of thrown objects and calculate the flight path of a target. By utilizing the initial velocity, throwing angle, and gravitational acceleration, the parabolic model can accurately predict the future trajectory of an object. In airport security, the parabolic motion model is used to fit the detected trajectory of a projectile and, combined with Kalman filtering, to correct the trajectory and accurately predict the target's landing point, thereby assessing potential threats.

[0035] Kalman filtering, a recursive algorithm for estimating the state of dynamic systems, is widely used in navigation, positioning, and target tracking. By combining a model and observations, Kalman filtering recursively updates the estimated target state, making it suitable for accurate state estimation in noisy environments. The basic process of Kalman filtering includes two stages: prediction and update. In the prediction stage, the current state is predicted based on the system's state transition model; in the update stage, the prediction result is corrected using current observations to obtain a more accurate estimate. In this invention, Kalman filtering is applied to the dynamic tracking of small targets, particularly in the trajectory tracking of high-speed flying targets. It can update the target's position and velocity in real time and improve the accuracy of the target trajectory by correcting prediction errors. In projectile detection, Kalman filtering, combined with a parabolic motion model, can effectively correct the target trajectory and predict the future target position, thereby improving the accuracy of projectile trajectory recognition and impact point prediction.

[0036] In step S4, the edge computer performs local fusion analysis on the results of the vaulting behavior recognition and the projectile trajectory recognition to construct a multimodal event fusion model. This avoids data interaction delays across devices. By making unified decisions based on behavior category, timestamp, spatial coordinates, trajectory information, and target screenshots, it determines whether the current event belongs to vaulting intrusion, projectile intrusion, or a combined composite threat. Once the edge computer determines that vaulting or projectile behavior exists, it immediately triggers an alarm and simultaneously generates standardized alarm data, reducing the cloud response stage. Specifically, when a dangerous behavior is determined, standardized alarm data is immediately generated, compressed in real time by the edge computer, and uploaded to the airport security system to reduce bandwidth consumption. The standardized alarm data includes event type, occurrence time, fence location coordinates, compressed screenshots of keyframes of the detected target, and multi-target tracking IDs. The standardized alarm data is controlled within 2KB.

[0037] Using high-definition infrared surveillance cameras at 30fps around the airport perimeter and an edge computer within 50 meters of the cameras (the edge computer is equipped with an NVIDIA Jetson AGX Orin GPU acceleration module), the system detects a pedestrian climbing over the fence and simultaneously throwing an object at a distance of 50 meters. The edge computer performs spatiotemporal behavior analysis and trajectory recognition in parallel, generating two alarm messages locally with an average detection latency of less than 250ms, and uploading standardized alarm data of less than 2KB to the airport security system.

[0038] By enabling data fusion and decision-making at the edge, cross-device information exchange and cloud processing latency are reduced, keeping alarm response time in the millisecond range. At the same time, frequent communication between cameras and the cloud is avoided, improving security efficiency and reducing cloud computing pressure.

[0039] In step S5, the edge computer pushes standardized alarm data to the airport security system in real time and can directly link perimeter defense hardware according to preset strategies for rapid response, including automatically turning on searchlights to illuminate suspected intrusion areas, triggering audible and visual alarms, and activating drone patrol systems or public address systems. In the cloud-based optional collaborative mode, alarm information is also simultaneously sent to the security backend for event archiving, dispatching, and subsequent review. Step S5 improves the immediacy, reliability, and automation level of airport perimeter protection, achieving efficient security management.

[0040] Example 2: like Figure 2 As shown in the figure, this embodiment provides an airport pedestrian fence-climbing and object-throwing detection and recognition system based on edge computing, including: Data preprocessing module 1 utilizes the ViBe or KNN algorithm to establish a stable background model, effectively eliminating background noise such as changes in illumination and swaying leaves. Simultaneously, it employs the DeepSORT algorithm to continuously track detected moving targets, generating a continuous image sequence with target bounding boxes and trajectory annotations. This module significantly reduces video transmission bandwidth consumption while outputting clear, structured target trajectory data, providing high-quality input for subsequent behavior recognition and trajectory analysis, and improving the system's real-time performance and data processing efficiency.

[0041] Spatiotemporal behavior analysis module 2 calls the human keypoint detection network to extract human skeleton keypoints, and connects the skeleton points of each frame to form a skeleton temporal sequence through time series analysis. The ST-GCN spatiotemporal graph convolutional network is then used to perform joint spatial topology and temporal dynamic modeling of the skeleton sequence, enabling the analysis of key points near the fence. Climbing Accurate identification across continuous actions. This module reduces the false positive rate of single-frame actions while supporting real-time inference, outputting a complete vaulting action sequence including behavior category, timestamp, and target location, providing reliable data for subsequent alarms and safety decisions.

[0042] The trajectory analysis and recognition module 3 uses the YOLOv8 small target detection model to identify high-speed flying objects. It uses Kalman filtering to track the target's position in consecutive frames and combines a parabolic motion model to fit the trajectory and predict the future landing point. It outputs the target's motion trajectory, velocity, acceleration, and landing point position, which can effectively distinguish between high-risk projectiles and natural disturbances, such as birds and falling leaves, providing accurate basis for real-time early warning while reducing false alarm rate and response delay.

[0043] The fusion decision processing module 4 performs parallel fusion of the trespassing behavior analysis results and the parabolic trajectory recognition results. By uniformly processing behavior categories, timestamps, spatial coordinates, trajectory information, and target screenshots, this module can determine the event type as trespassing intrusion, parabolic intrusion, or a combined threat, and generate standardized alarm data. This module localizes the decision-making process, achieving millisecond-level response, reducing cross-device and cloud communication latency, and improving the accuracy of threat identification and system reliability.

[0044] The early warning and linkage execution module 5 pushes the generated standardized alarm data to the airport security system in real time and drives the perimeter defense hardware to perform immediate linkage according to the strategy, including automatically turning on searchlights, triggering audible and visual sirens, launching drone patrols, or broadcasting announcements. This module ensures that a complete closed-loop response can be completed even in weak network or network outage environments, enabling rapid threat handling. Through structured data transmission and automated execution, the module comprehensively improves the immediacy, reliability, and automation level of airport perimeter security management.

[0045] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The methods disclosed in the embodiments are described simply because they correspond to the systems disclosed in the embodiments; relevant details can be found in the method section.

[0046] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0047] In the embodiments provided by this invention, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0048] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0049] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit.

[0050] Similarly, in the various embodiments of the present invention, each processing unit can be integrated into a functional module, or each processing unit can exist physically, or two or more processing units can be integrated into a functional module.

[0051] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0052] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0053] The above-disclosed embodiments are merely preferred embodiments of the present invention, but the present invention is not limited thereto. Any non-creative variations that can be conceived by those skilled in the art, as well as any improvements and modifications made without departing from the principles of the present invention, should fall within the protection scope of the present invention.

Claims

1. A method for detecting and recognizing pedestrians climbing over fences and projectiles at airports based on edge computing, characterized in that, Includes the following steps: Step S1: The data preprocessing step involves real-time acquisition of video streams from high-definition surveillance cameras around the airport, local background modeling and moving target detection by the edge computer, and running a multi-target tracking algorithm. Step S2: The step of spatiotemporal behavior analysis is to use the human key point detection network to extract the human skeleton point sequence and construct a spatiotemporal graph convolutional network to identify continuous behavior patterns of climbing and traversing. Step S3: The trajectory analysis and identification step involves the edge computer extracting high-speed small targets and calculating their trajectories based on a trajectory fitting algorithm, distinguishing high-risk projectiles and estimating their landing points. Step S4: The step of fusion decision processing, the edge computer performs joint analysis on the climbing behavior recognition result and the parabolic trajectory recognition result locally, triggers alarm and generates standardized alarm data; Step S5: The step of early warning linkage execution. The edge computer pushes the compressed alarm information to the airport security system in real time and links with the perimeter defense hardware.

2. The method for detecting and recognizing pedestrians climbing over fences and projectiles at airports based on edge computing, as described in claim 1, is characterized in that... In step S1: video streams are acquired in real time from high-definition surveillance cameras at the airport perimeter and transmitted to an edge computer at the airport perimeter. The edge computer uses ViBe or KNN background modeling algorithms to establish a stable dynamic background model, extracts motion regions based on foreground segmentation results, and uses target detection algorithms to detect moving targets. At the same time, the edge computer runs the DeepSORT multi-target tracking algorithm to continuously track moving targets. After the above processing, the original video stream is converted into a continuous image sequence containing foreground segmentation results, target detection boxes, and stable target trajectory annotations.

3. A method for detecting and recognizing pedestrians climbing over fences and projectiles at airports based on edge computing, as described in claim 1 or 2, characterized in that... In step S2: an edge computer is used to perform spatiotemporal behavior analysis on the continuous image sequence obtained in step S1. A human keypoint detection network is called to extract key skeleton points of multiple parts of the human body to form a frame-level skeleton structure. These skeletons are then connected into a human skeleton temporal sequence through the time dimension. The action sequence is modeled using the ST-GCN spatiotemporal graph convolutional network to construct a spatiotemporal graph model based on key points. The human action is jointly convolved in both spatial topology and temporal dynamics to achieve accurate recognition of continuous vaulting behavior patterns and output the vaulting behavior recognition result.

4. The method for detecting and recognizing pedestrians climbing over fences and projectiles at airports based on edge computing, as described in claim 3, is characterized in that... The result of the trespassing behavior recognition is a precisely identified sequence of trespassing behaviors and related attributes including timestamps, target locations, and behavior categories.

5. The method for detecting and recognizing pedestrians climbing over fences and projectiles at airports based on edge computing, as described in claim 4, is characterized in that... In step S3: the YOLOv8 small target detection model is used to extract small targets from the continuous image sequence obtained in step S1, extracting high-speed flying projectiles, including bottles, stones, and drones. Kalman filtering is used to predict the position of the detected small targets at the next moment. Subsequently, the time series position points of the small targets are input into the parabolic motion model and combined with Kalman filtering for trajectory fitting. The velocity, acceleration, and direction of the object's flight path are estimated, high-risk projectiles are distinguished from naturally disturbed targets, and their future landing positions are predicted. The output is the trajectory recognition result of the projectile, including the target's motion trajectory, velocity, acceleration, and landing position.

6. The method for detecting and recognizing pedestrians climbing over fences and projectiles at airports based on edge computing, as described in claim 5, is characterized in that... In step S4: the edge computer performs fusion analysis on the results of the vaulting behavior recognition and the trajectory recognition of the projectile locally to construct a multimodal event fusion model. By making unified decisions on behavior category, timestamp, spatial coordinates, trajectory information and target screenshot, it determines whether the current event belongs to vaulting intrusion, projectile intrusion or a combined composite threat. If the edge computer determines that there is vaulting or projectile behavior, it immediately triggers an alarm and generates standardized alarm data simultaneously. After being compressed in real time by the edge computer, the information is uploaded to the airport security system.

7. The method for detecting and recognizing pedestrians climbing over fences and projectiles at airports based on edge computing as described in claim 6, characterized in that, The standardized alarm data includes event type, occurrence time, fence location coordinates, compressed screenshot of keyframe of detected target, and multi-target tracking ID. The standardized alarm data is controlled within 2KB.

8. The method for detecting and recognizing pedestrians climbing over fences and projectiles at airports based on edge computing, as described in claim 7, is characterized in that... In step S5, the edge computer pushes standardized alarm data to the airport security system in real time and can directly link the perimeter defense hardware for rapid response according to preset strategies, including automatically turning on searchlights to illuminate suspected intrusion areas, triggering sound and light alarms, starting drone patrol systems or broadcasting devices. In the cloud-selectable collaborative mode, alarm information will also be sent to the security backend simultaneously.

9. An airport pedestrian fence-climbing and object-throwing detection and recognition system based on edge computing, characterized in that, include: Data preprocessing module (1), spatiotemporal behavior analysis module (2), trajectory analysis and recognition module (3), fusion decision processing module (4), early warning linkage execution module (5); The data preprocessing module (1) is used to collect video streams in real time from high-definition surveillance cameras at the airport perimeter and perform background modeling, moving target detection and multi-target tracking locally on the edge computer; The spatiotemporal behavior analysis module (2) is used to extract the human skeleton point sequence through the human key point detection network, and to use the spatiotemporal graph convolutional network to model and analyze, and to identify the behavior of pedestrians climbing over fences. The trajectory analysis and recognition module (3) is used to fit the trajectory of a high-speed small target in the video, identify parabolic behavior and predict its landing area; The fusion decision processing module (4) is used to fuse the identification results of climbing behavior and parabolic behavior, perform joint analysis and trigger an alarm; The early warning linkage execution module (5) is used to push the compressed alarm information to the airport security system in real time and drive the perimeter defense hardware to link together.

10. The airport pedestrian fence crossing and projectile detection and recognition system based on edge computing according to claim 9, characterized in that, The data preprocessing module (1) uses the ViBe or KNN algorithm to establish a stable background model, uses the DeepSORT algorithm to continuously track the detected moving targets, and generates a continuous image sequence with target bounding boxes and trajectory annotations. The spatiotemporal behavior analysis module (2) calls the human key point detection network to extract human skeleton key points, and connects each frame skeleton point through time series to form skeleton time sequence. It uses ST-GCN spatiotemporal graph convolutional network to perform spatial topology and temporal dynamic joint modeling of skeleton sequence to accurately identify continuous actions. The trajectory analysis and recognition module (3) uses the YOLOv8 small target detection model to identify high-speed flying objects, uses Kalman filtering to track the position of the target in continuous frames, and combines the parabolic motion model to fit the trajectory and predict the future landing point, outputting the target's motion trajectory, speed, acceleration and landing point position; The fusion decision processing module (4) performs parallel fusion of the overcrowding behavior analysis results and the parabolic trajectory recognition results. By uniformly processing the behavior category, timestamp, spatial coordinates, trajectory information and target screenshot, it determines the event type as overcrowding intrusion, parabolic intrusion or composite threat, and generates standardized alarm data. The early warning linkage execution module (5) pushes the generated standardized alarm data to the airport security system in real time, and drives the perimeter defense hardware to perform real-time linkage according to the strategy, including automatically turning on the searchlight, triggering the sound and light alarm, starting the drone patrol or broadcasting.

Images