Machine vision-based antitheft system for cultural relics display cases and identification method

Abstract

The application provides a machine vision-based anti-climbing early warning system for cultural relic display cabinets and a recognition method, and belongs to the technical field of visual detection, and comprises the following steps: acquiring a monitoring video stream of an area where a cultural relic display cabinet is located and physical size data of the cultural relic display cabinet; extracting a human body skeleton key point coordinate sequence of an audience by using a pre-trained human body skeleton key point detection model; calculating a posture jump index and a spatial approximation degree; calculating a height-cabinet matching risk coefficient; analyzing and acquiring a climbing behavior coherence degree; calculating a comprehensive climbing risk index; and triggering a corresponding anti-climbing early warning action when the comprehensive climbing risk index exceeds a preset threshold. The application solves the problem that the prior art is difficult to fuse the physical size of the display cabinet and the spatial posture of the audience to accurately predict and grade the early warning of the climbing behavior.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology, specifically to a machine vision-based anti-climbing early warning system and identification method for cultural relic display cases. Background Technology

[0002] With increasingly stringent security requirements in museums, art galleries, and other venues, the security of artifact display cases has expanded from simply preventing theft and vandalism to proactively issuing warnings about various violations by visitors. Climbing display cases, as a destructive and dangerous activity, can not only cause cases to tip over and artifacts to be damaged, but also easily lead to public safety accidents such as personal injury. Therefore, it is a high-risk scenario that requires close attention in exhibition security.

[0003] Existing anti-climbing technologies for display cases mainly fall into three categories. The first category is physical barriers, which increase the difficulty of climbing by setting up barriers, protective railings, or using sloping tops around the display case. While this provides some passive protection, it disrupts the overall visual effect of the exhibition and cannot prevent climbing over barriers or using raised platforms. The second category is detection methods based on single sensors, such as using infrared beams, pressure-sensitive mats, or ultrasonic ranging to detect approaching targets. These methods generally suffer from limited detection capabilities, making it difficult to distinguish between normal close-up viewing and genuine climbing intentions, resulting in a high false alarm rate. The third category is detection methods based on traditional video analysis, which typically rely solely on changes in the target detection frame's position for simple judgments, ignoring the subtle changes in human posture and the interaction between the display case's dimensions, leading to insufficient accuracy in behavior recognition and timeliness of warnings.

[0004] In summary, existing technologies lack an early warning mechanism that can accurately distinguish between normal viewing behavior and gradual climbing intentions, making it difficult to achieve highly reliable real-time early warnings with a low false alarm rate. Summary of the Invention

[0005] This invention provides a machine vision-based anti-climbing early warning system and identification method for cultural relic display cases, which solves the problem that existing technologies are unable to accurately predict and provide graded early warnings of climbing behavior by integrating the physical dimensions of the display case with the spatial posture of the audience.

[0006] In view of the above problems, the present invention provides a machine vision-based anti-climbing early warning system and identification method for cultural relic display cases.

[0007] In a first aspect, the present invention provides a machine vision-based anti-climbing early warning system for cultural relic display cases, comprising: The data acquisition module acquires the monitoring video stream of the area where the cultural relic display case is located and the physical dimension data of the cultural relic display case. The physical dimension data includes at least the height value of the top surface of the display case and the horizontal reference position of the outer side of the display case. The human skeleton key point extraction module uses a pre-trained human skeleton key point detection model to extract the coordinate sequence of human skeleton key points of the audience, wherein the human skeleton key points include at least the center point of the hip joint and the center points of both ankle joints. The posture and spatial relationship calculation module calculates the posture rise index based on the relative relationship between the vertical coordinate of the center point of the hip joint and the height value of the top surface of the display case, and calculates the spatial approximation based on the distance relationship between the horizontal coordinate of the center point of the two ankle joints and the horizontal reference position of the outer side of the display case. The height-showcase matching risk coefficient calculation module calculates the height-showcase matching risk coefficient by combining the estimated height of the audience with the height of the top surface of the showcase. The climbing behavior coherence analysis module analyzes and obtains the climbing behavior coherence based on the changing trends of the attitude leap index and the spatial approximation in consecutive frames. The comprehensive risk assessment module calculates a comprehensive climbing risk index based on the posture leap index, the spatial proximity, the height-display case matching risk coefficient, and the climbing behavior continuity. The early warning trigger module triggers a corresponding anti-climbing early warning action when the comprehensive climbing risk index exceeds a preset threshold.

[0008] Secondly, this invention provides a machine vision-based anti-climbing early warning and identification method for cultural relic display cases, including: Acquire the surveillance video stream of the area where the cultural relic display case is located and the physical dimension data of the cultural relic display case, wherein the physical dimension data includes at least the height value of the top surface of the display case and the horizontal reference position of the outer side of the display case; A pre-trained human skeleton key point detection model is used to extract the coordinate sequence of human skeleton key points of the audience, wherein the human skeleton key points include at least the center point of the hip joint and the center points of both ankle joints. Based on the relative relationship between the vertical coordinates of the hip joint center point and the height of the top surface of the display case, the posture rise index is calculated, and based on the distance relationship between the horizontal coordinates of the center points of the two ankle joints and the horizontal reference position of the outer side of the display case, the spatial approximation is calculated. The height-showcase matching risk coefficient is calculated by combining the estimated audience height with the height of the top surface of the display case. Based on the changing trends of the attitude leap index and the spatial approximation in consecutive frames, the coherence of climbing behavior is analyzed and obtained. A comprehensive climbing risk index is calculated based on the posture leap index, the spatial proximity, the height-display case matching risk coefficient, and the climbing behavior coherence. When the comprehensive climbing risk index exceeds a preset threshold, the corresponding anti-climbing warning action is triggered.

[0009] One or more technical solutions provided in this invention have at least the following technical effects or advantages: The technical solution of this invention first acquires real-time surveillance video streams of the area where the artifact display case is located using monitoring equipment. Based on the display case number, it queries a pre-stored database for the height value of the top surface of the display case and the horizontal reference position of the outer side of the display case, unifying visual input and spatial geometric parameters under the same spatiotemporal coordinate system. This provides a precise and standardized spatial reference benchmark for all subsequent analyses, enabling the determination of climbing behavior to be based on an objective measurement of the three-dimensional spatial relationship between the viewer's body and the display case.

[0010] Furthermore, a pre-trained skeletal keypoint detection model is used to detect and track spectators appearing in the surveillance video stream frame by frame, outputting 3D coordinate sequences of key skeletal points such as the center points of the hip joint and both ankle joints, along with age classification labels. This transforms the spatial behavior of the spectators into quantifiable and trackable skeletal point coordinate data, providing a high-precision foundation for subsequent posture analysis.

[0011] Furthermore, the posture elevation index and spatial proximity are calculated simultaneously to quantify the upward shift of the viewer's center of gravity relative to the top of the display case and the closeness of their feet to the edge of the display case, respectively. This creates a three-dimensional spatial posture description from both vertical and horizontal dimensions, accurately capturing the core precursory features of climbing behavior and effectively compensating for the high false alarm rate and delayed judgment of single indicators.

[0012] Furthermore, by utilizing the stable vertical distance between the top of the head and the center points of both ankle joints, and after verification with age classification labels, an estimated height of the audience is obtained. This height is then compared with the height of the top surface of the display case to calculate the height-display case matching risk coefficient. This method identifies individuals with natural climbing abilities from a static physical characteristic perspective, providing differentiated prior weights for risk assessment and avoiding invalid alerts for short individuals without climbing ability.

[0013] Furthermore, the temporal trends of attitude leap index and spatial approximation are analyzed within a preset frame window. The coherence of climbing behavior is obtained by calculating the total change, net change, and directional consistency coefficient frame by frame. This effectively distinguishes between continuous climbing intentions and occasional attitude fluctuations from a temporal perspective, significantly reducing false alarms caused by non-climbing behaviors such as tiptoeing to view or bending over to tie shoelaces.

[0014] Furthermore, a weighted fusion of postural leap index, spatial proximity, height-display case matching risk coefficient, and climbing behavior consistency is performed, with weights configured differently based on audience age classification labels to calculate a comprehensive climbing risk index. This approach integrates multi-dimensional heterogeneous features into a unified decision criterion, taking into account the differences in athletic ability among audiences of different age groups, and resolving the issues of one-sided evaluation dimensions and inconsistent judgment standards associated with single indicators.

[0015] Finally, based on a dual-channel mechanism combining manual confirmation and automatic system labeling to accumulate historical audience behavior samples, dynamic and updatable first and second risk thresholds are set. The comprehensive climbing risk index is compared with the two thresholds in a graded manner, triggering a first-level warning with silent monitoring and voice prompts, or a second-level warning with voice warnings combined with human intervention. A complete closed loop from risk quantification to graded response is established, gently dissuading climbers at the initial stage and forcibly intervening when the danger escalates, achieving refined and tiered display case security protection.

[0016] In summary, the technical solution of the present invention achieves active identification, early prediction and differentiated response to climbing behavior without visually interfering with the exhibition effect or changing the original structure of the display case. It effectively solves the core problems of high false alarm rate, delayed judgment and insufficient adaptability to different age groups in the prior art. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the anti-climbing early warning system for cultural relic display cases based on machine vision provided by the present invention.

[0018] Figure 2 This is a schematic diagram illustrating the calculation process of the comprehensive climbing risk index in the anti-climbing early warning and identification method for cultural relic display cases based on machine vision provided by the present invention.

[0019] Figure 3 This is a flowchart illustrating the anti-climbing early warning and identification method for cultural relic display cases based on machine vision provided by the present invention.

[0020] In the attached diagram, the labels representing each component are as follows: Data acquisition module 11, human skeleton key point extraction module 12, posture and spatial relationship calculation module 13, height-display case matching risk coefficient calculation module 14, climbing behavior coherence analysis module 15, comprehensive risk assessment module 16, early warning triggering module 17. Detailed Implementation

[0021] This invention provides a machine vision-based anti-climbing early warning system and identification method for cultural relic display cases, which solves the problem that existing technologies are unable to accurately predict and provide graded early warnings of climbing behavior by integrating the physical dimensions of the display case with the spatial posture of the audience.

[0022] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0023] Example 1, as Figure 1 As shown, this invention provides a machine vision-based anti-climbing early warning system for cultural relic display cases, the system comprising: The data acquisition module 11 acquires the monitoring video stream of the area where the cultural relic display case is located and the physical dimension data of the cultural relic display case. The physical dimension data includes at least the height value of the top surface of the display case and the horizontal reference position of the outer side of the display case.

[0024] This module is used to acquire the surveillance video stream of the area where the artifact display case is located and to simultaneously load the physical dimension data of the artifact display case. The physical dimension data includes at least the height of the top surface of the display case and the horizontal reference position of the outer side of the display case. The former provides a vertical spatial reference for determining whether the viewer's center of gravity exceeds the top of the display case, and the latter provides a horizontal distance reference for measuring how close the viewer's feet are to the edge of the display case.

[0025] The steps of module 11 in the system provided by this invention include: The monitoring equipment was used to collect surveillance video streams of the area where the artifact display case was located; Based on the display case number, the top height of the display case and the horizontal reference position of the outer side of the display case are obtained as the physical dimension data. The top height of the display case is the vertical distance of the top of the display case relative to the floor of the exhibition hall, and the horizontal reference position of the outer side of the display case is the horizontal coordinate value of the outer side of the display case facing the audience in the coordinate system of the monitoring screen.

[0026] In this step, the first step is to use surveillance equipment deployed in the area where the artifact display cases are located to collect real-time video streams of that area. For example, directly in front of the bronze tripod display case, a high-definition surveillance camera is mounted at a downward angle on the ceiling of the exhibition hall to continuously collect real-time video streams of the area in front of the display case, covering the front of the display case and the area where visitors are active.

[0027] Furthermore, based on the display case number of the cultural relic display case, the corresponding physical dimension data is retrieved from the pre-stored display case information database, specifically including the height value of the top surface of the display case and the horizontal reference position of the outer side of the display case. The height value of the top surface of the display case is the vertical distance between the top surface of the cultural relic display case and the floor plane of the exhibition hall, and the horizontal reference position of the outer side of the display case is the horizontal coordinate value of the outer side of the cultural relic display case facing the audience in the coordinate system of the monitoring screen.

[0028] For example, based on the display case number ZT-003, its physical dimension parameters are retrieved from the pre-stored display case information database: the vertical distance from the top surface of the display case to the floor of the exhibition hall is 1.15 meters, and the horizontal reference line coordinate of the outer edge of the front of the display case in the coordinate system of the monitoring screen is Y=420 pixels.

[0029] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0030] In summary, this step provides a precise and standardized spatial reference benchmark for subsequent quantitative calculations of the attitude leap amplitude and spatial approximation degree based on the key points of the audience's skeleton. This enables the determination of climbing behavior to be based on an objective measurement of the three-dimensional spatial relationship between the audience's body and the display case, ensuring the accuracy of the early warning system's identification from the data source.

[0031] The human skeleton key point extraction module 12 uses a pre-trained human skeleton key point detection model to extract the coordinate sequence of human skeleton key points of the audience, wherein the human skeleton key points include at least the center point of the hip joint and the center points of both ankle joints.

[0032] This module employs a pre-trained human skeletal keypoint detection model to perform frame-by-frame inference on viewers appearing in the surveillance video stream, extracting and outputting a sequence of human skeletal keypoint coordinates. The extracted keypoints include at least the center point of the hip joint and the center points of both ankle joints, which are used to represent the viewer's center of gravity position and lower limb stance, respectively.

[0033] The steps of module 12 in the system provided by this invention include: The monitoring video stream is input frame by frame into a pre-trained human skeletal key point detection model. Human detection and skeletal key point localization are performed on each of the audience members appearing in each frame. The human skeletal key point coordinate sequence and corresponding age classification label of each audience member with the highest age confidence are output. The human skeletal key points include at least the three-dimensional spatial coordinates of the center point of the hip joint and the center points of both ankle joints. The skeletal key points of the same audience member in multiple consecutive frames are tracked and associated to form a sequence of human skeletal key point coordinates of the same audience member in the time dimension.

[0034] In this step, the pre-training of the human skeletal keypoint detection model includes: Collect a dataset of sample images containing human postures of visitors in museum exhibition hall scenes, and annotate the human skeletal key points and age classification labels in each sample image to form a sample annotation set. The skeletal key points include at least the center point of the hip joint and the center points of both ankle joints of the sample. The sample image dataset is divided into child sample subsets, adult sample subsets, and elderly sample subsets according to human body shape characteristics and mobility ability; A skeletal keypoint detection model is constructed, comprising an input layer, a shared feature extraction layer, three parallel age branch layers, and an output layer. The three parallel age branch layers include a child branch, an adult branch, and an elderly branch. The skeletal keypoint detection model is jointly trained using the sample image dataset and the sample annotation set until convergence, thus obtaining the trained skeletal keypoint detection model.

[0035] Specifically, we first collected a dataset of sample images containing human postures of visitors in museum exhibition hall scenes. We then labeled the key points of the human skeleton and age classification labels in each sample image to form a sample annotation set. Among them, the key points of the skeleton include at least the center point of the hip joint and the center points of both ankle joints of the sample.

[0036] When collecting sample image datasets from museum exhibition scenes, the resolution of the sample images was uniformly normalized to 1280×720 pixels to match the mainstream output resolution of surveillance cameras. The sample annotation set annotated each image with 17 key points of the human skeleton, including the center points of the hip joints and both ankle joints, as well as the top of the head, neck, both shoulder joints, both elbow joints, both wrist joints, and both knee joints. Age classification labels were divided into age ranges: children 3 to 12 years old, adults 18 to 60 years old, and the elderly 60 years and older.

[0037] For example, in the exhibition hall where the bronze tripod is displayed, surveillance cameras captured video footage of visitors over a period of 30 days. From this footage, 5000 sample images were extracted, showing visitors of different body types and ages in various postures, including standing, bending over, tiptoeing, and raising their hands. For each image, 17 key skeletal points, such as the center point of the hip joint and the center points of both ankle joints, were individually labeled, along with an age classification tag. Specifically, a sample image of a child visitor, with a resolution of 1280×720 pixels, was labeled with the pixel coordinates of its 17 key skeletal points, and also labeled as a child.

[0038] Secondly, the sample image dataset was divided into child, adult, and elderly subsets based on body shape characteristics and motor ability. The division was based on height range and skeletal proportion parameters. The child subset corresponds to a height range of 0.8 to 1.5 meters, with a head-to-body ratio of approximately 1:5 to 1:6; the adult subset corresponds to a height range of 1.5 to 1.9 meters, with a head-to-body ratio of approximately 1:7 to 1:8; and the elderly subset corresponds to a height range of 1.4 to 1.8 meters, but incorporates a kyphosis angle parameter, with a spinal lordosis angle ranging from 5 to 25 degrees.

[0039] For example, based on age classification labels, 5,000 sample images were divided into three subsets: approximately 800 images of children with shorter stature and different bone proportions, approximately 3,200 images of adults with significant differences in body size and strong motor skills, and approximately 1,000 images of elderly people with relatively slow posture characteristics. Specifically, a child audience member with a height of approximately 1.05 meters and a head-to-body ratio of approximately 1:5.5 was included in the children's sample subset; an adult audience member with a height of approximately 1.72 meters and a head-to-body ratio of approximately 1:7.5 was included in the adult sample subset; and an elderly audience member with a height of approximately 1.65 meters and a spinal lordosis angle of approximately 15 degrees was included in the elderly sample subset.

[0040] Next, a skeletal keypoint detection model is constructed, comprising an input layer, a shared feature extraction layer, three parallel age branch layers, and an output layer. The three parallel age branch layers include a children's branch, an adult's branch, and an elderly branch, each specifically learning the skeletal proportions and pose features of its corresponding age group.

[0041] In the constructed skeletal keypoint detection model, the shared feature extraction layer uses ResNet-50 as the backbone network, with an input image size of 640×480 pixels and an output feature map size of 20×15×2048. Three parallel age branch layers each contain a fully connected layer and a regression layer. The fully connected layers each have 1024 neurons, using ReLU as the activation function and a Dropout ratio of 0.5 to prevent overfitting. Each branch's regression layer outputs 34 values, corresponding to the two-dimensional pixel coordinates of 17 skeletal keypoints, each represented as (X,Y). After obtaining the two-dimensional pixel coordinates, combined with the intrinsic and extrinsic parameter calibration matrix of the surveillance camera, a monocular visual geometric transformation is used to map the two-dimensional pixel coordinates to a three-dimensional spatial coordinate system based on the exhibition hall floor plane, obtaining the three-dimensional spatial coordinates (X,Y,Z) of each keypoint. Here, the Z coordinate is the vertical height of the keypoint relative to the exhibition hall floor plane, and the X and Y coordinates are the horizontal projection positions of the keypoints on the exhibition hall floor plane. The initial learning rate for each branch was set to 0.001, and the weight decay coefficient was set to 0.0005.

[0042] The camera's intrinsic and extrinsic parameter calibration matrices are pre-acquired during system deployment. The Zhang Zhengyou calibration method is employed, using a checkerboard calibration board of known dimensions to capture images at multiple positions and angles within the display area. This process determines the camera's intrinsic parameters, such as focal length and principal point coordinates, as well as extrinsic parameters, such as the camera's rotation matrix and translation vector relative to the exhibition hall's floor coordinate system. The calibration process is a one-time operation; once completed, the parameters are stored in the system configuration file for real-time conversion from 2D pixel coordinates to 3D spatial coordinates during the inference phase.

[0043] The conversion from pixel coordinates to physical coordinates is performed in the following order after each inference: First, the two-dimensional pixel coordinates output by the model are combined with the camera intrinsic parameter matrix to correct lens distortion and map to normalized imaging plane coordinates; second, using the rotation matrix and translation vector in the camera extrinsic parameters, combined with the plane equation constraints of the exhibition hall floor plane, the normalized coordinates are projected onto the exhibition hall floor plane, and the physical coordinates (X, Y) of each keypoint on the exhibition hall floor plane are solved, where the Z coordinate is the vertical height of the keypoint relative to the exhibition hall floor plane. The conversion process is performed in real time frame by frame to ensure that the skeletal keypoint coordinates output in each frame are three-dimensional coordinates in physical space.

[0044] For example, after a viewer image with an input size of 640×480 pixels is entered into the model, it is extracted into a 20×15×2048 feature map by the ResNet-50 shared layer, and then fed into three parallel branches. The child branch is processed by a 1024-dimensional fully connected layer and outputs 34 key point coordinate prediction values.

[0045] Finally, the skeletal keypoint detection model was jointly trained using the sample image dataset and the sample annotation set until the model converged, resulting in a trained skeletal keypoint detection model. During the joint training phase, the Adam optimizer was used with a batch size of 32 and an initial learning rate of 0.001, which decreased by 0.1 times every 50 training epochs. The loss function consisted of two weighted components: the keypoint coordinate regression loss used mean squared error with a weight of 0.7; the age classification loss used cross-entropy loss with a weight of 0.3. The total number of training epochs was set to 200. An early stopping mechanism was triggered when the validation set loss did not decrease for 10 consecutive epochs. The model convergence threshold was set when the validation loss fell below 0.02.

[0046] For example, sample images from three subsets and their corresponding annotations are input into the model in batches. The keypoint detection loss and age classification loss of each branch are jointly calculated, and the parameters of the shared layer and each branch are iteratively updated through backpropagation. During joint training, each sample image is simultaneously forward-inferred through the three branches, and the keypoint regression loss of each branch is calculated. The true age classification label of the sample is used as a supervision signal to calculate the age classification loss. After the keypoint regression loss and age classification loss of each branch are weighted and summed, the specific parameters of the shared feature extraction layer and each branch are simultaneously updated through backpropagation. The above training mechanism enables the shared layer to learn the general spatial distribution features of skeletal points, while each branch refines the skeletal proportions and pose features of the corresponding age group. Ultimately, this achieves a reasoning paradigm of three parallel inference paths under the same input, with the highest confidence level as the output.

[0047] During training, the training loss was 2.35 in the first round, decreasing to 0.18 in the 100th round, 0.05 in the 150th round, and 0.018 in the 180th round, falling below the convergence threshold of 0.02, triggering early stopping and ending training. At this point, the average keypoint detection errors for the child branch, adult branch, and elderly branch were 1.8 pixels, 1.5 pixels, and 2.1 pixels, respectively. The model was saved as a pre-trained skeletal keypoint detection model that could be deployed on-site at the bronze tripod display case.

[0048] The pre-trained human skeletal key point detection model refers to a deep learning network that has been trained in advance through the above process and is capable of accurately extracting skeletal key points from visitors of different ages in a museum exhibition setting. When deployed to an actual exhibition case monitoring system, it can be directly loaded and used without the need for on-site retraining or annotation.

[0049] In this step, the video stream is first input frame by frame into a pre-trained human skeletal keypoint detection model. Each frame is simultaneously processed through a shared feature extraction layer and then fed into separate branches for children, adults, and the elderly for parallel inference. Each branch independently outputs the predicted coordinates of the viewer's human skeletal keypoints and their corresponding confidence scores. The keypoint coordinate sequence output by the branch with the highest confidence score, along with its corresponding age classification label, is taken as the preliminary result. Subsequently, using a pre-calibrated camera intrinsic and extrinsic parameter matrix, the two-dimensional pixel coordinates are converted into three-dimensional spatial coordinates based on the exhibition hall floor plane. Finally, the three-dimensional spatial coordinate sequence of each skeletal keypoint and its age classification label are output. The human skeletal keypoints include at least the three-dimensional spatial coordinates of the hip joint center point and the center points of both ankle joints.

[0050] For example, in the monitoring screen of the bronze tripod display case, the current frame shows both an adult visitor and a child visitor. For the first visitor, the confidence level of the child branch output is 0.12, the confidence level of the adult branch output is 0.95, and the confidence level of the elderly branch output is 0.08. Taking the output result of the adult branch with the highest confidence level, the two-dimensional pixel coordinates of the center point of the hip joint are (620, 380), which, after conversion by camera calibration parameters, yields three-dimensional spatial coordinates of (1.85, 0.42, 0.98) (unit: meters), and the age classification label is adult. For the second visitor, the confidence level of the child branch output is 0.92, the confidence level of the adult branch output is 0.31, and the confidence level of the elderly branch output is 0.15. Taking the output result of the child branch with the highest confidence level, the two-dimensional pixel coordinates of the center point of the hip joint are (320, 450), which, after conversion by camera calibration parameters, yields three-dimensional spatial coordinates of (0.95, 0.38, 0.68), and the age classification label is child.

[0051] Furthermore, the skeletal key points of the same viewer in multiple consecutive frames are tracked and correlated to form a sequence of human skeletal key point coordinates of the same viewer in the time dimension.

[0052] For example, in the subsequent 50 consecutive frames, the skeletal key points of the adult viewer are continuously tracked using Kalman filtering and Hungarian matching algorithms. The three-dimensional spatial coordinates of the hip joint center point and the center points of both ankle joints, obtained after extracting the two-dimensional pixel coordinates and converting them by camera calibration in each frame, are concatenated into a temporal sequence. For example, the Z-coordinate of the hip joint in frame 1 is 0.98 meters, in frame 10 it is 0.99 meters, in frame 20 it is 1.05 meters, and in frame 50 it is 1.20 meters, forming a complete dynamic trajectory sequence describing the gradual upward movement of the viewer's center of gravity.

[0053] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0054] In summary, this step uses the center point of the hip joint as a spatial representation of the viewer's body center of gravity and the center points of both ankle joints as a precise description of the viewer's standing and movement. It transforms the viewer's spatial behavior in front of the display case into quantifiable and traceable skeletal point coordinate data, providing a high-precision, high-temporal-resolution foundation for subsequent calculations of posture rise index and spatial approximation. This enables the early identification of climbing behavior to achieve a technological leap from coarse target detection to fine posture analysis.

[0055] The posture and spatial relationship calculation module 13 calculates the posture rise index based on the relative relationship between the vertical coordinate of the center point of the hip joint and the height value of the top surface of the display case, and calculates the spatial approximation based on the distance relationship between the horizontal coordinate of the center point of the two ankle joints and the horizontal reference position of the outer side of the display case.

[0056] This module performs two quantitative calculations simultaneously. On the one hand, based on the relative relationship between the vertical coordinates of the viewer's hip joint center point and the height of the top surface of the display case, it calculates the postural leap index, which is used to quantify the upward shift of the viewer's center of gravity relative to the top surface of the display case. On the other hand, based on the distance relationship between the horizontal coordinates of the viewer's ankle joint center points and the horizontal reference position on the outer side of the display case, it calculates the spatial proximity, which is used to quantify the closeness of the viewer's lower limbs to the edge of the display case.

[0057] The steps of module 13 in the system provided by this invention include: The arithmetic mean of the vertical coordinates of the hip joint center point within a preset initial number of frames after the audience enters the monitoring screen is obtained as the reference height of the hip joint when standing normally. The deviation between the current frame's hip joint height value and the normal standing hip joint reference height is calculated and used as the posture rise index. Obtain the horizontal coordinates of the midpoint of the center point of the ankle joints of the two feet in the current frame, and use them as the horizontal position of the two feet; Calculate the shortest Euclidean distance between the horizontal position of the two feet and the horizontal reference position of the outer side of the display case, and combine it with the warning distance of the display case to calculate and obtain the spatial approximation.

[0058] In this step, the arithmetic mean of the vertical coordinates of the hip joint center point within a preset initial frame number after the viewer enters the monitoring screen is first obtained, which is used as the reference height of the hip joint for normal standing. The preset initial frame number refers to the first few frames when the viewer just enters the monitoring screen and has not yet shown any climbing intention. It is usually the first 10 to 15 frames after the viewer is detected, which is used as the reference benchmark for the viewer's normal standing posture.

[0059] For example, after an adult visitor enters the monitoring screen of the bronze tripod display case, the vertical coordinates of the center point of the hip joint in the first 10 frames are taken, which are 0.98 meters, 0.99 meters, 0.97 meters, 1.00 meters, 0.98 meters, 0.99 meters, 0.96 meters, 0.98 meters, 1.00 meters, and 0.97 meters respectively. The arithmetic mean is calculated to be 0.982 meters, which is taken as the reference height of the visitor's normal standing hip joint.

[0060] Furthermore, the deviation between the current frame's hip height value and the baseline hip height for normal standing is calculated and used as the posture rise index. The posture rise index is calculated as follows: the difference between the current frame's hip height value and the baseline height is divided by the baseline height to obtain a normalized deviation ratio. When the posture rise index approaches 0, it indicates that the viewer's hip height is at the same level as in a normal standing state; a positive index indicates that the body's center of gravity has shifted upwards relative to a normal standing state, with a larger value indicating a greater upward shift; a negative index indicates that the body's center of gravity has shifted upwards relative to a normal standing state, such as when the viewer squats or bends over.

[0061] Since the core characteristic of climbing behavior in the vertical direction is the upward shift of the body's center of gravity, and actions such as squatting or bending down that involve a downward shift of the center of gravity contradict the directionality of climbing, negative values ​​of the posture rise index are truncated: when the posture rise index is negative, it is set to 0, indicating that the current audience has not exhibited the upward shift of the center of gravity required for climbing in the vertical direction. The rationale for this is that climbing behavior requires the audience to simultaneously meet two conditions: an upward shift of the center of gravity and close proximity to the display case. A downward shift of the center of gravity itself does not constitute a threat to climbing and does not need to be included in the subsequent weighted calculation of negative contributions, but negative values ​​are still retained in the original data records for behavioral retrospective analysis.

[0062] For example, when the audience tiptoed and supported themselves on the edge of the display case with their hands, their hip joint height rose to 1.20 meters, which is 0.218 meters higher than the baseline height of 0.982 meters. The postural rise index = 0.218 / 0.982 ≈ 0.222, indicating that their center of gravity has shifted significantly upward.

[0063] In another possible embodiment, if a visitor bends down to tie their shoelaces in front of the display case, their current frame hip height is 0.50 meters, which differs from the baseline height of 0.982 meters by -0.482 meters. The original posture rise index is approximately -0.482 / 0.982 -0.491. After truncation, the posture rise index is set to 0, indicating that the visitor is not currently exhibiting an upward shift of their center of gravity, and the climbing risk assessment relies solely on other indicators such as spatial proximity and height matching risk coefficient.

[0064] Furthermore, obtain the horizontal coordinates of the midpoint of the center points of both ankle joints in the current frame, as the horizontal position of both feet. The horizontal position of both feet is the geometric midpoint between the horizontal coordinates of the center points of the left and right ankle joints.

[0065] For example, if the horizontal coordinates of the center point of the left ankle joint in the current frame are (600, 580) and the horizontal coordinates of the center point of the right ankle joint are (640, 580), then the midpoint coordinates of the horizontal positions of both feet are (620, 580).

[0066] Finally, the pixel coordinates of the horizontal position of the feet are converted into physical coordinates. Using the pre-calibrated camera intrinsic and extrinsic parameter matrices and the mapping relationship of the exhibition hall floor plane, the pixel coordinates (620, 580) are converted into physical coordinates on the exhibition hall floor plane. At the same time, the horizontal reference position of the outer side of the display case is stored in the form of pixel coordinates Y=420, which is also converted into physical coordinates through the camera calibration parameters, representing the horizontal position line of the outer side of the display case facing the audience on the exhibition hall floor plane.

[0067] Since the horizontal reference position on the outer side of the display case is a horizontal straight line along the width of the display case in physical space, parallel to the floor of the exhibition hall, the distance from the viewer's feet to the reference line is the distance from the physical position of the feet to the straight line in the direction perpendicular to the outer side of the display case. The physical distance from the feet to the horizontal reference position on the outer side of the display case can be obtained by calculating the absolute value of the difference between the component of the physical coordinates of the feet in this direction and the corresponding component of the reference line.

[0068] The warning distance for display cases is a preset safety buffer distance threshold, typically ranging from 0.8 meters to 1.5 meters. The specific value is determined based on the spatial layout of the exhibition hall, visitor density, and the security level of the exhibits. For independent display cases with wide aisles, the warning distance should be set at 1.2 to 1.5 meters; for display cases against a wall or in narrow aisles, the warning distance should be set at 0.8 to 1.0 meters; for first-class cultural relics or high-security exhibits, the warning distance can be appropriately increased to 1.5 meters to provide more sufficient warning buffer time.

[0069] For example, after calibration and transformation, the horizontal reference position on the outer side of the bronze tripod display case has a depth-direction coordinate of 3.50 meters in the physical coordinate system of the exhibition hall floor. The pixel coordinates (620, 580) of the viewer's horizontal feet, after calibration and transformation, have a depth-direction coordinate of 3.35 meters in the physical coordinate system of the exhibition hall floor. The physical distance from the viewer's feet to the horizontal reference position on the outer side of the display case is |3.35 - 3.50| = 0.15 meters. The preset warning distance for the display case is 1.5 meters, and the spatial proximity is 1 - (0.15 / 1.5) = 0.90, indicating that the viewer is very close to the display case.

[0070] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0071] In summary, this step, by simultaneously calculating the posture elevation index and spatial proximity, achieves two key benefits. Firstly, the posture elevation index accurately captures the core precursor to climbing behavior—the rise of the viewer's center of gravity—eliminating interference from simple standing or slight swaying. Secondly, the spatial proximity objectively measures the changing distance between the viewer and the display case, distinguishing between long-distance viewing and close-range interaction. The combination of these two indicators provides a three-dimensional quantitative description of the viewer's spatial posture in both vertical and horizontal directions, offering reliable data support from two dimensions for early prediction of climbing intentions. This effectively compensates for the high false alarm rate and delayed judgment associated with single indicators.

[0072] The height-showcase matching risk coefficient calculation module 14 calculates the height-showcase matching risk coefficient by combining the estimated height of the audience with the height of the top surface of the showcase.

[0073] This module calculates the height-showcase matching risk coefficient by combining the estimated height of the viewer with the height of the top of the display case. The estimated height of the viewer is obtained based on the stable value of the vertical distance from the top of the head to the center point of both ankle joints in multi-frame skeletal keypoint data, and is verified in conjunction with age classification labels. The height-showcase matching risk coefficient is used to assess whether the viewer's physical condition provides an innate advantage in easily climbing the display case. It remains constant after the viewer first enters the monitoring screen and completes the height estimation, serving as a static physical indicator to supplement the risk assessment from a physical dimension, and does not fluctuate due to subsequent changes in the viewer's standing posture.

[0074] The steps of module 14 in the system provided by this invention include: The stable value of the vertical distance between the head vertex and the center point of the two ankle joints in the skeletal key points is obtained within a preset stable frame number after the audience enters the monitoring screen. Combined with the age classification label output by the human skeletal key point detection model, the audience's height is estimated. The ratio of the estimated audience height to the top height of the display case is calculated to obtain the height-display case matching risk coefficient.

[0075] In this step, the vertical distance between the top of the viewer's head and the center points of both ankles is first obtained within the first few frames after the viewer enters the monitoring screen. This vertical distance is taken as the mode or median of the measurements from multiple frames when the viewer is standing normally and has not yet shown any intention to climb, to eliminate interference from accidental tiptoeing or bending over. Specifically, the stable value of the difference between the Z-coordinate of the top of the head and the Z-coordinate of the center points of both ankles within the first 15 to 30 frames is directly used as the viewer's estimated height. Combined with age classification labels, the reasonableness of this estimated value is verified: if the estimated value exceeds ±20% of the typical height range for that age group, the sampling frame number is extended to obtain a new stable value.

[0076] For example, in front of the bronze tripod display case, after an adult visitor enters the monitoring screen, the stable value of the vertical distance between the top of the head and the center points of both ankle joints is taken from the first 20 frames. The first 5 data points for each frame are 1.83 meters, 1.84 meters, 1.83 meters, 1.85 meters, and 1.83 meters. Among them, 1.83 meters appears most frequently and is within the typical height range of adults, i.e., 1.5 meters to 1.9 meters. The stable value of 1.83 meters is taken as the estimated height of the visitor. According to the age classification label corresponding to the branch with the highest confidence after the three-way parallel inference of the model, the branch with the highest confidence of the visitor being an adult has passed the verification, and the estimated height is determined to be 1.83 meters.

[0077] Furthermore, the ratio of the estimated audience height to the top height of the display case is calculated to obtain the height-display case matching risk coefficient. This coefficient is calculated for the first time after the audience enters the frame and is then locked as a constant, not updated with subsequent changes in the audience's posture, ensuring its consistency as a static physical indicator. The specific calculation method for the height-display case matching risk coefficient is as follows: the estimated audience height is divided by the top height of the display case to obtain a ratio. When this ratio is greater than or equal to 1.5, it indicates that the audience's height is more than 1.5 times the height of the display case, possessing a significant climbing advantage, and the height-display case matching risk coefficient is set to its maximum value of 1.0. When this ratio is less than or equal to 0.5, it indicates that the audience's height is less than half the height of the display case, making independent climbing almost impossible, and the risk coefficient is set to its minimum value of 0.0. When the ratio is between 0.5 and 1.5, the risk coefficient is linearly mapped to the interval [0,1].

[0078] For example, if the estimated height of the visitor is 1.83 meters and the top height of the bronze tripod display case is 1.15 meters, the ratio of the estimated height to the top height is approximately 1.59 (1.83 / 1.15). This ratio exceeds 1.5, resulting in a height-display case matching risk coefficient of 1.0, indicating that the visitor's height makes climbing extremely easy. If we consider a child visitor with an estimated height of only 0.90 meters, the ratio would be approximately 0.78 (0.90 / 1.15), which is greater than 0.5 but less than 1.5. Therefore, the linear mapping yields a height-display case matching risk coefficient of 0.28.

[0079] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0080] In summary, this step, starting from the static physical characteristics of the audience, quantifies the objective matching relationship between height and display case height into a continuously calculable indicator. It can automatically identify and focus on individuals whose height conditions provide natural climbing convenience, providing differentiated prior weights for subsequent comprehensive risk assessment. Thus, without increasing hardware costs, it effectively improves the targeting and accuracy of climbing warnings and avoids generating invalid alarms for short people who cannot climb.

[0081] The climbing behavior coherence analysis module 15 analyzes and obtains the climbing behavior coherence based on the changing trends of the attitude leap index and the spatial approximation in consecutive frames.

[0082] This module analyzes and obtains the coherence of climbing behavior based on the changing trends of the attitude rise index and spatial approximation over multiple consecutive frames. By examining whether the two spatial pose indices show a synergistic increasing correlation in the time dimension, it distinguishes between sporadic isolated attitude fluctuations and intentionally deliberate, continuously progressive sequences of climbing actions.

[0083] The steps of module 15 in the system provided by this invention include: Obtain the attitude leap index sequence and spatial approximation sequence of the audience within a preset frame window; Calculate the absolute value of the attitude rise index difference between each two adjacent frames in the attitude rise index sequence, and sum them to obtain the total change in attitude rise frame by frame. Calculate the absolute value of the spatial approximation difference between each two adjacent frames in the spatial approximation sequence, and sum them to obtain the total change in spatial approximation frame by frame; Calculate the maximum difference in attitude rise index in the attitude rise index sequence within a preset frame window as the net change in attitude rise, and calculate the maximum difference in spatial approximation in the spatial approximation sequence within a preset frame window as the net change in spatial approximation. Calculate the ratio of the net change in attitude rise to the total change in attitude rise frame by frame to obtain the attitude rise direction consistency factor, and calculate the ratio of the net change in spatial approximation to the total change in spatial approximation frame by frame to obtain the spatial approximation direction consistency factor. Calculate the arithmetic mean of the attitude jump direction consistency factor and the spatial approximation direction consistency factor to obtain the direction consistency coefficient; The product of the directional consistency coefficient and the larger of the net change in attitude elevation and the net change in spatial approximation is used as the climbing behavior coherence.

[0084] In this step, the audience's attitude rise index sequence and spatial approximation sequence within a preset frame window are first obtained. The preset frame window employs a frame-by-frame sliding mechanism. Specifically, with each frame arriving in the monitored video stream, the current frame is used as the end point of the window, and a preset number of frames (e.g., 30 frames) are traced back. The attitude rise index sequence and spatial approximation sequence within this sliding window are extracted, and a climbing behavior coherence calculation is performed. When the next frame arrives, the window simultaneously slides forward one frame, and the latest 30-frame sequence is extracted again for calculation.

[0085] In contrast, if a non-overlapping skip-analysis is performed every fixed number of frames, the initial climbing action within the interval between two analysis sessions may be completely missed, and decision delays are likely to occur at the skip boundaries. Therefore, this invention employs a frame-by-frame sliding window mechanism to obtain the highest temporal sensitivity and full coverage of behavioral changes in continuous video streams.

[0086] For example, in front of the bronze tripod display case, a preset frame window of 30 frames is used, employing a frame-by-frame sliding mechanism. When the video stream reaches frame 100, the attitude rise index sequence and spatial approximation sequence of 30 frames (frames 71 to 100) are extracted for coherence analysis. When the video stream reaches frame 101, the window automatically slides to frames 72 to 101, and the analysis is performed again. This frame-by-frame approach generates a coherence evaluation value for each frame, ensuring that the instantaneous behavioral change of the viewer starting to tiptoe and climb around frame 76 can be captured immediately. The following are examples of sequences within the window from frame 71 to frame 100: An example of the attitude rise exponent sequence is [0.06, 0.07, 0.08, 0.13, 0.20, 0.26, 0.33, 0.40, 0.48, 0.53, 0.56, 0.58, 0.58, 0.60, 0.62, 0.63, 0.61, 0.66, 0.68, 0.68, 0.69, 0.70, 0.70, 0.71, 0.72, 0.72, 0.73, 0.73]. [0.73, 0.76]; an example of the spatial approximation sequence is [0.10, 0.10, 0.11, 0.13, 0.17, 0.20, 0.25, 0.30, 0.35, 0.39, 0.42, 0.45, 0.48, 0.51, 0.53, 0.55, 0.57, 0.58, 0.59, 0.60, 0.61, 0.61, 0.62, 0.62, 0.63, 0.63, 0.63, 0.63, 0.63].

[0087] Furthermore, the absolute value of the difference in attitude rise index between each two adjacent frames in the attitude rise index sequence is calculated and summed to obtain the total change in attitude rise per frame; at the same time, the absolute value of the difference in spatial approximation between each two adjacent frames in the spatial approximation sequence is calculated and summed to obtain the total change in spatial approximation per frame.

[0088] For example, for the above attitude rise index sequence, the absolute value of the difference between adjacent frames is calculated and summed. The total change in attitude rise per frame = 0.01 + 0.01 + 0.05 + 0.07 + 0.06 + 0.07 + 0.07 + 0.08 + 0.05 + 0.03 + 0.02 + 0.00 + 0.02 + 0.02 + 0.01 + 0.02 + 0.05 + 0.02 + 0.00 + 0.01 + 0.01 + 0.00 + 0.01 + 0.01 + 0.00 + 0.01 + 0.00 + 0.00 + 0.03 = 0.73. Similarly, the total change in spatial approximation frame by frame = 0.00 + 0.01 + 0.02 + 0.04 + 0.03 + 0.05 + 0.05 + 0.05 + 0.04 + 0.03 + 0.03 + 0.03 + 0.02 + 0.02 + 0.02 + 0.01 + 0.01 + 0.01 + 0.00 + 0.01 + 0.00 + 0.00 + 0.00 + 0.00 + 0.00 = 0.53.

[0089] Furthermore, the maximum difference in attitude rise index in the attitude rise index sequence within the preset frame window is calculated as the net change in attitude rise, and the maximum difference in spatial approximation in the spatial approximation sequence within the preset frame window is calculated as the net change in spatial approximation.

[0090] For example, in the attitude rise index sequence, the maximum value is 0.76 and the minimum value is 0.06. The net change in attitude rise is 0.76 - 0.06 = 0.70. The truncation process ensures that the net change is always non-negative, and the calculation of the orientation consistency factor is not affected by negative values. In the spatial approximation sequence, the maximum value is 0.63 and the minimum value is 0.10. The net change in spatial approximation is 0.63 - 0.10 = 0.53.

[0091] Specifically, the net change in attitude elevation is compared with a preset net change threshold. If the net change in attitude elevation is less than the preset net change threshold, the climbing behavior continuity of that window is directly set to 0, and subsequent directional consistency factor and continuity calculations are skipped. At the same time, the net change in spatial approximation is compared with a preset spatial approximation net change threshold. If the net change in spatial approximation is less than the preset spatial approximation net change threshold, the climbing behavior continuity of that window is directly set to 0.

[0092] The preset net change threshold is set between 0.15 and 0.30, with a typical value of 0.20. When the net change in posture elevation is less than 0.20, it is considered that the change in posture within the window is insufficient to constitute climbing attention, and no further analysis is required. The preset net change threshold for spatial approach is set between 0.10 and 0.20, with a typical value of 0.15. When the net change in spatial approach is less than 0.15, it is considered that the audience has not engaged in effective distance reduction behavior within the window.

[0093] The net change threshold is set based on the following: the threshold must be greater than the fluctuation range of the posture rise index generated by the natural swaying of the body when the audience is standing normally, usually not exceeding 0.10, and less than the increase in the posture rise index of mild tiptoeing behavior, usually about 0.15 to 0.25.

[0094] For example, the above-mentioned attitude jump net change amount = 0.70, which is greater than the preset net change amount threshold of 0.20, and spatial approximation net change amount = 0.53, which is greater than the preset spatial approximation net change amount threshold of 0.15, both pass the validity judgment and proceed to the subsequent calculation of the directional consistency factor.

[0095] Furthermore, the ratio of the net change in attitude rise to the total change in attitude rise per frame is calculated to obtain the attitude rise direction consistency factor; and the ratio of the net change in spatial approximation to the total change in spatial approximation per frame is calculated to obtain the spatial approximation direction consistency factor.

[0096] For example, the attitude rise direction consistency factor is calculated to be approximately 0.96 (0.70 / 0.73), indicating that the attitude rise index generally increases within the window, with only a slight pullback around frame 17, maintaining a high level of directional consistency. The spatial approximation direction consistency factor is approximately 1.0 (0.53 / 0.53 + 0.80 / 0.80), indicating that the spatial approximation degree increases strictly and monotonically within the window, without any backward fluctuations.

[0097] Furthermore, the arithmetic mean of the attitude rise direction consistency factor and the spatial approximation direction consistency factor is calculated to obtain the direction consistency coefficient.

[0098] For example, the directional consistency coefficient is calculated as (0.96 + 1.0) / 2 = 0.98, indicating that the audience exhibits a highly consistent unidirectional approach trend in both the vertical and horizontal directions.

[0099] Finally, the directional consistency coefficient is multiplied by the larger of the net change in attitude jump and the net change in spatial approximation, and the resulting product is taken as the climbing behavior coherence.

[0100] For example, the larger of the net change in posture elevation (0.70) and the net change in spatial approximation (0.53) is 0.70. The climbing behavior coherence is 0.98 × 0.70 = 0.69, indicating that the spectator's behavior within the 1-second time window has extremely high coherence and a clear climbing intention.

[0101] In another possible embodiment, if the audience only briefly tiptoes before standing up and backing away, the posture leap index will first rise and then fall, and the directional consistency coefficient may be only 0.3. The coherence of the climbing behavior will be greatly reduced, and no subsequent warning will be triggered.

[0102] In another possible embodiment, another viewer stands close to the bronze tripod display case and bends down to tie his shoelaces. Within a 30-frame window, the spatial proximity continuously increases from 0.10 to 0.63, with a net change in spatial proximity of 0.53. However, the viewer's center of gravity continues to descend, and the posture rise index sequence is all 0 after being truncated due to negative values. The net change in posture rise is 0, which is less than the preset net change threshold of 0.20. The continuity of the climbing behavior is directly set to 0, correctly excluding the above behavior.

[0103] In another possible embodiment, if a visitor repeatedly jumps in front of the display case, the attitude jump index sequence within a 30-frame window is [0.10, 0.35, 0.60, 0.75, 0.546, 0.20]. The maximum value is 0.76, the minimum value is 0.08, and the net attitude jump change is 0.68, which is greater than the net change threshold of 0.20, so it proceeds to the next step. The total change per frame is 2.92, and the attitude jump direction consistency factor is 0.68 / 2.92≈0.23. The spatial approximation sequence is [0.30, 0.45, 0.20, 0.50, 0.25, 0.48, 0.22, 0.52, 0.28, 0.45, 0.20, 0.48, 0.25, 0.50, 0.22, 0.55, 0.30, 0.48, 0.25, 0.50, 0.28, 0.55, 0.32, 0.48, 0.26, 0.50, 0.30, 0.52, 0.28, 0.48]. The maximum value is 0.55, the minimum value is 0.20, and the net change in spatial approximation is 0.35. Total frame-by-frame variation = 0.15 + 0.25 + 0.30 + 0.25 + 0.23 + 0.26 + 0.30 + 0.24 + 0.17 + 0.25 + 0.28 + 0.23 + 0.25 + 0.28 + 0.33 + 0.25 + 0.18 + 0.23 + 0.25 + 0.22 + 0.27 + 0.23 + 0.16 + 0.22 + 0.24 + 0.20 + 0.22 + 0.24 + 0.20 = 3.73. Spatial approximation directional consistency factor = 0.35 / 3.73 ≈ 0.09. Directional consistency coefficient = (0.23 + 0.09) / 2 = 0.16. Of the net change in attitude elevation (0.68) and the net change in spatial approach (0.35), the larger value is 0.68. Therefore, the climbing behavior consistency is approximately 0.11 (0.16 × 0.68). Although the net change exceeds the threshold, the repeated jumps result in a low directional consistency factor, and the consistency remains low. Further comprehensive risk assessment can incorporate other indicators such as spatial approach to effectively reduce the risk of false alarms.

[0104] In another possible embodiment, if a viewer stands normally in front of the display case and sways their body slightly, the maximum value of the attitude jump index within a 30-frame window is 0.12 and the minimum value is 0.05. The net change in attitude jump is 0.07, which is less than the preset net change threshold of 0.20. The climbing behavior continuity is directly set to 0, and there is no need to continue calculating the direction consistency factor. This behavior is correctly excluded from the climbing risk assessment.

[0105] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0106] In summary, the climbing behavior continuity calculated in this step can effectively distinguish between continuous, progressive movements with a clear climbing intention and occasional, brief postural fluctuations from a time dimension. By incorporating the consistency of movement direction and the cumulative amplitude into the quantitative assessment, the intentional characteristics can be identified before the climbing behavior has completed its full movement. This provides a highly discriminative dynamic behavioral basis for comprehensive risk assessment and significantly reduces the false alarm rate caused by non-climbing behaviors such as spectators tiptoeing to watch or briefly jumping to take photos.

[0107] The comprehensive risk assessment module 16 calculates a comprehensive climbing risk index based on the posture leap index, the spatial proximity, the height-display case matching risk coefficient, and the climbing behavior continuity.

[0108] like Figure 2 As shown, this module calculates a comprehensive climbing risk index based on four quantitative indicators: posture leap index, spatial proximity, height-display case matching risk coefficient, and climbing behavior coherence. By weighting and fusing instantaneous spatial posture characteristics, static physical risk, and dynamic behavioral coherence, a unified numerical value that comprehensively reflects the current climbing risk level of the audience is generated.

[0109] The steps of module 16 in the system provided by this invention include: The posture leap index, spatial proximity, height-display case matching risk coefficient, and climbing behavior coherence are weighted and calculated to obtain a comprehensive climbing risk index. The weighting factor is obtained based on the age classification label of the audience.

[0110] This step obtains the comprehensive climbing risk index through weighted calculation. Specifically, it involves a weighted sum of four indicators: posture leap index, spatial proximity, height-display case matching risk coefficient, and climbing behavior consistency. The weighting units are obtained based on the age classification labels of the audience. Each weighting unit consists of four preset weight values ​​for different age groups, corresponding to the four indicators mentioned above, with the sum of each weight set being 1.0.

[0111] The following are examples of weighting rules: For adults, who have stronger motor skills and can climb quickly, the weight of dynamic behavioral indicators should be higher than that of static physical characteristics indicators. A typical weighting configuration is: Postural leap index weight 0.25, Spatial approach weight 0.20, Height-display case matching risk coefficient weight 0.15, Climbing behavior continuity weight 0.40. For children, who are naturally active but have limited climbing ability, the weight of static physical characteristics indicators should be appropriately increased to differentiate between height differences. A typical weighting configuration is: Postural leap index weight 0.20, Spatial approach weight 0.20, Height-display case matching risk coefficient weight 0.15, Climbing behavior continuity weight 0.40. The height-display case matching risk coefficient is weighted at 0.35, and the climbing behavior continuity weight is weighted at 0.25. This means that the weight of dynamic continuity is appropriately reduced, while the reference weight of the height matching risk coefficient is increased. For the elderly, whose movements are slow and climbing continuity is usually low, the weight of climbing behavior continuity should be reduced, while the weights of spatial approach and postural leap index are increased to capture the slow but continuous approach trend. A typical weight configuration is: postural leap index weight 0.30, spatial approach weight 0.30, height-display case matching risk coefficient weight 0.20, and climbing behavior continuity weight 0.20.

[0112] For example, in front of the bronze tripod display case, based on the age classification label corresponding to the branch with the highest confidence after three-way parallel inference of the model, the branch indicating that the visitor is an adult has the highest confidence, and the corresponding weighted reassembly is invoked. The current values ​​of each indicator are: Postural leap index 0.76, Spatial approximation 0.63, Height-display case matching risk coefficient 1.0, Climbing behavior coherence 0.69. The comprehensive climbing risk index = 0.25 × 0.76 + 0.20 × 0.63 + 0.15 × 1.0 + 0.40 × 0.69 = 0.742.

[0113] Since the comprehensive climbing risk index is calculated based on a frame-by-frame sliding window, an updated risk index can be obtained for each frame. The warning triggering module performs a threshold comparison once for each frame to ensure that the delay between the risk index crossing the threshold and the triggering of the warning action does not exceed the time interval of a single frame, thus achieving real-time response at the frame level.

[0114] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0115] In summary, this step integrates the assessment results from four dimensions—vertical body center of gravity elevation, horizontal standing proximity, static physical characteristics (height convenience), and temporal behavioral intent consistency—into a unified and comparable comprehensive risk quantification index. This achieves an effective mapping from multi-source heterogeneous characteristics to a single decision criterion, ensuring the comprehensiveness of risk assessment and providing a clear and age-appropriate judgment benchmark for subsequent early warning triggering. It fundamentally solves the problems of one-sided single-indicator assessment dimensions and inconsistent risk judgment standards for audiences of different age groups.

[0116] The early warning trigger module 17 triggers a corresponding anti-climbing early warning action when the comprehensive climbing risk index exceeds a preset threshold.

[0117] This module compares the overall climbing risk index with a preset threshold in real time. When the overall climbing risk index exceeds the preset threshold, a corresponding anti-climbing warning action is triggered. The preset threshold is a configurable risk assessment threshold. The anti-climbing warning action is divided into different levels according to the extent to which the overall climbing risk index exceeds the limit, and may include local audio and visual warnings, remote security center pop-up alarms, on-site broadcast voice reminders, and other multi-level response measures.

[0118] The steps of module 17 in the system provided by this invention include: Based on historical audience behavior samples, a first risk threshold and a second risk threshold are obtained, wherein the first risk threshold is smaller than the second risk threshold; When the comprehensive climbing risk index is less than the first risk threshold, it is determined that the audience is within the range of normal visiting behavior, and silent monitoring is maintained. When the comprehensive climbing risk index is greater than or equal to the first risk threshold and less than the second risk threshold, a level one warning is triggered, wherein the level one warning includes at least a voice broadcast prompt; When the comprehensive climbing risk index is greater than or equal to the second risk threshold, a level-two warning is triggered, wherein the level-two warning includes at least a voice broadcast prompt and the issuance of a human intervention dispatch instruction.

[0119] In this step, a first risk threshold and a second risk threshold are first obtained based on historical audience behavior samples. The first risk threshold is lower than the second risk threshold, and the two together constitute the judgment interval for tiered early warning. Historical audience behavior samples refer to the comprehensive risk index statistical set of normal visitor behavior data and verified climbing behavior data accumulated during the system's deployment and operation. The two types of samples are obtained through the following methods: Normal visitor behavior samples are automatically collected by the system. When a visitor's comprehensive climbing risk index remains below the first risk threshold for a continuously preset duration without triggering any warning, the comprehensive climbing risk index value of each frame within that period is automatically included in the normal visitor behavior sample set. The continuously preset duration is typically 30 seconds to ensure sufficient stability of the collected behavioral data.

[0120] Confirmed climbing behavior samples are obtained through a dual-channel confirmation mechanism. Channel 1 is manual backend confirmation: When a level 2 alert is triggered, security personnel verify the alert event via handheld terminals or the monitoring center backend. If the event is confirmed as genuine climbing behavior, the comprehensive climbing risk index value of all frames within the event period is automatically added to the confirmed climbing behavior sample set. If it is confirmed as a false alarm, the corresponding data is added to the normal visitor behavior sample set for subsequent threshold calibration. Channel 2 is auxiliary automatic confirmation: When the comprehensive climbing risk index continuously exceeds a preset multiple of the second risk threshold for a preset duration, and both the attitude jump index and spatial approximation exceed preset high thresholds, the event is automatically marked as high-confidence climbing behavior and directly added to the confirmed climbing behavior sample set without manual confirmation, thus reducing the burden of manual labeling. The preset duration is typically 5 seconds, the preset multiple is 1.2 times, and the preset high threshold is 0.85.

[0121] The threshold calculation rules include: the first risk threshold is usually set at the 95th percentile of the comprehensive risk index of historical normal visitor behavior samples to cover the vast majority of normal behavior fluctuations; the second risk threshold is usually set at the 5th percentile of the comprehensive risk index of historical confirmed climbing behavior samples to ensure complete coverage of actual climbing behavior. Using the 5th percentile instead of the minimum value avoids the second threshold being too low due to individual low-risk abnormal samples, thus enhancing the statistical robustness of the thresholds.

[0122] The thresholds employ a dynamic update mechanism, meaning the specific values ​​of both thresholds are dynamically updated as the system continuously accumulates samples. A periodic batch update strategy can be used, automatically triggering a threshold recalculation whenever the cumulative number of newly added normal visitor behavior samples reaches a preset quantity or at preset time intervals. A cooldown period is set to prevent frequent threshold jumps. The preset quantity is typically 5000 samples, the preset time period is 7 days, and the cooldown period is no less than 24 hours.

[0123] The threshold update incorporates a smoothing mechanism, calculated as follows: First risk threshold update value = 95th percentile of the normal sample in the current period × update smoothing coefficient + First risk threshold of the previous period × (1 - update smoothing coefficient); Second risk threshold update value = 5th percentile of the climbing sample in the current period × update smoothing coefficient + Second risk threshold of the previous period × (1 - update smoothing coefficient). The update smoothing coefficient ranges from 0.3 to 0.7, with a typical value of 0.5, to balance the weights of old and new statistical values ​​and avoid significant threshold fluctuations due to statistical bias in a single batch of samples.

[0124] If, after the update, the first risk threshold is greater than or equal to the second risk threshold, then the second risk threshold will be forcibly set to equal the first risk threshold plus the minimum safety interval. The minimum safety interval is set to 0.5 times the standard deviation of historically confirmed climbing behavior samples, and not less than 0.05, to ensure that a reasonable buffer zone is always maintained between the two thresholds.

[0125] In the initial deployment phase, before a sufficient number of historical samples have been accumulated, the first risk threshold is initially set to 0.40, and the second risk threshold is initially set to 0.65. Once 1,000 samples of normal visitor behavior and 20 confirmed samples of climbing behavior have been accumulated, the thresholds will be updated for the first time using the aforementioned dynamic calculation model.

[0126] For example, in the initial deployment of the bronze tripod display case monitoring system, the initial risk thresholds were set as follows: the first risk threshold was 0.40, and the second risk threshold was 0.65. After three months of operation, 12,000 samples of normal visitor behavior were accumulated, with the 95th percentile at 0.45; 50 samples of confirmed climbing behavior were manually confirmed, with the 5th percentile at 0.68. Using an update smoothing coefficient of 0.5, the update calculations were: First risk threshold = 0.45 × 0.5 + 0.40 × 0.5 = 0.425; Second risk threshold = 0.68 × 0.5 + 0.65 × 0.5 = 0.665. The update was activated when the constraint that the first risk threshold was less than the second risk threshold was met, and the area between the two thresholds constituted the observation interval for potential climbing risk. Thereafter, an update was triggered every 7 days or every additional 5,000 normal samples, with the threshold values ​​gradually approaching the true distribution as data accumulated.

[0127] Secondly, when the overall climbing risk index is less than the first risk threshold, it is determined that the audience is within the range of normal visiting behavior, and silent monitoring is maintained without triggering any alarm actions.

[0128] For example, if a visitor stands normally 1.8 meters in front of the bronze tripod display case, their comprehensive climbing risk index is calculated to be 0.18, which is lower than the first risk threshold of 0.45. This is considered normal visiting behavior, and the visitor is kept under silent monitoring.

[0129] Furthermore, a Level 1 warning is triggered when the comprehensive climbing risk index is greater than or equal to the first risk threshold but less than the second risk threshold. A Level 1 warning includes at least a voice announcement to remind visitors to observe proper etiquette and maintain a safe distance, serving as a preventative measure to discourage potential incidents.

[0130] For example, if an adult visitor begins to approach the display case and tiptoe, their posture elevation index rises to 0.50, spatial proximity rises to 0.40, height-display case matching risk coefficient is 1.0, and climbing behavior continuity is 0.45. The comprehensive climbing risk index is calculated as follows: 0.25 × 0.50 + 0.20 × 0.40 + 0.15 × 1.0 + 0.40 × 0.45 = 0.535. Since the comprehensive climbing risk index is greater than or equal to the first risk threshold of 0.45 and less than the second risk threshold of 0.68, a level one warning is triggered. The hidden speakers in the display case area automatically play a voice prompt: "Please maintain a safe distance, be courteous when viewing the exhibit, thank you for your cooperation."

[0131] Finally, when the comprehensive climbing risk index is greater than or equal to the second risk threshold, a level-two warning is triggered. A level-two warning includes at least a voice broadcast prompt and the issuance of human intervention and dispatch instructions. At this point, it is determined that the climbing behavior has a high degree of real danger, and security personnel must be dispatched to the scene immediately for handling.

[0132] For example, the aforementioned adult visitor ignored the Level 1 warning and continued climbing, causing the posture ascent index to rise to 0.76, the spatial proximity to reach 0.63, and the climbing behavior continuity to increase to 0.69. The comprehensive climbing risk index was recalculated to 0.742, which is greater than or equal to the second risk threshold of 0.68. The system immediately triggered a Level 2 warning, and the loudspeaker played a warning voice message: "Your behavior has triggered a safety alarm. Please stop immediately." At the same time, a human intervention dispatch instruction was automatically generated and pushed to the handheld terminal of the security personnel closest to the bronze tripod display case. The instruction included the display case number, location information, a real-time screenshot of the visitor, and a risk level label.

[0133] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0134] In summary, this step establishes a direct mapping closed loop from multi-dimensional risk assessment to tiered early warning response. It can automatically trigger voice prompts for gentle dissuasion in the early stages of climbing behavior, and immediately issue human intervention and dispatch instructions to enforce measures when the behavior escalates to a high degree of danger. This achieves accurate, timely, and differentiated response to climbing threats while ensuring that the visitor experience is not excessively disturbed. It fundamentally solves the problem that traditional security measures are passive and lagging behind, and cannot effectively intervene in the early stages of climbing behavior.

[0135] In summary, this invention achieves proactive identification, early prediction, and differentiated response to climbing behavior without visually interfering with the exhibition effect or altering the original structure of the display case. It effectively solves the core problems of existing technologies, such as high false alarm rate, delayed judgment, and insufficient adaptability to different age groups.

[0136] Example 2, as Figure 3As shown, this invention provides a machine vision-based anti-climbing warning and identification method for cultural relic display cases, the method comprising: S100: Obtain the surveillance video stream of the area where the cultural relic display case is located and the physical dimension data of the cultural relic display case, wherein the physical dimension data includes at least the height value of the top surface of the display case and the horizontal reference position of the outer side of the display case.

[0137] The acquisition of the surveillance video stream of the area where the artifact display case is located and the physical dimension data of the artifact display case includes: The monitoring equipment was used to collect surveillance video streams of the area where the artifact display case was located; Based on the display case number, the top height of the display case and the horizontal reference position of the outer side of the display case are obtained as the physical dimension data. The top height of the display case is the vertical distance of the top of the display case relative to the floor of the exhibition hall, and the horizontal reference position of the outer side of the display case is the horizontal coordinate value of the outer side of the display case facing the audience in the coordinate system of the monitoring screen.

[0138] S200: The pre-trained human skeleton key point detection model is used to extract the coordinate sequence of human skeleton key points of the audience, wherein the human skeleton key points include at least the center point of the hip joint and the center points of both ankle joints.

[0139] The method employs a pre-trained human skeletal keypoint detection model to extract the coordinate sequence of human skeletal keypoints from the viewer. These keypoints include at least the center point of the hip joint and the center points of both ankle joints. The monitoring video stream is input frame by frame into a pre-trained human skeletal key point detection model. Human detection and skeletal key point localization are performed on each of the audience members appearing in each frame. The human skeletal key point coordinate sequence and corresponding age classification label of each audience member with the highest age confidence are output. The human skeletal key points include at least the three-dimensional spatial coordinates of the center point of the hip joint and the center points of both ankle joints. The skeletal key points of the same audience member in multiple consecutive frames are tracked and associated to form a sequence of human skeletal key point coordinates of the same audience member in the time dimension.

[0140] The pre-training of the human skeletal key point detection model includes: Collect a dataset of sample images containing human postures of visitors in museum exhibition hall scenes, and annotate the human skeletal key points and age classification labels in each sample image to form a sample annotation set. The skeletal key points include at least the center point of the hip joint and the center points of both ankle joints of the sample. The sample image dataset is divided into child sample subsets, adult sample subsets, and elderly sample subsets according to human body shape characteristics and mobility ability; A skeletal keypoint detection model is constructed, comprising an input layer, a shared feature extraction layer, three parallel age branch layers, and an output layer. The three parallel age branch layers include a child branch, an adult branch, and an elderly branch. The skeletal keypoint detection model is jointly trained using the sample image dataset and the sample annotation set until convergence, thus obtaining the trained skeletal keypoint detection model.

[0141] S300: Based on the relative relationship between the vertical coordinates of the center point of the hip joint and the height of the top surface of the display case, calculate the posture rise index, and based on the distance relationship between the horizontal coordinates of the center points of the two ankle joints and the horizontal reference position of the outer side of the display case, calculate the spatial approximation.

[0142] The calculation of the posture rise index based on the relative relationship between the vertical coordinates of the hip joint center point and the height of the top surface of the display case, and the calculation of the spatial approximation based on the distance between the horizontal coordinates of the ankle joint center points and the horizontal reference position of the outer side of the display case, includes: The arithmetic mean of the vertical coordinates of the hip joint center point within a preset initial number of frames after the audience enters the monitoring screen is obtained as the reference height of the hip joint when standing normally. The deviation between the current frame's hip joint height value and the normal standing hip joint reference height is calculated and used as the posture rise index. Obtain the horizontal coordinates of the midpoint of the center point of the ankle joints of the two feet in the current frame, and use them as the horizontal position of the two feet; Calculate the shortest Euclidean distance between the horizontal position of the two feet and the horizontal reference position of the outer side of the display case, and combine it with the warning distance of the display case to calculate and obtain the spatial approximation.

[0143] S400: Calculate the height-showcase matching risk coefficient by combining the estimated audience height with the height of the top surface of the display case.

[0144] The calculation of the height-display case matching risk coefficient, which combines the estimated audience height with the height of the top surface of the display case, includes: The stable value of the vertical distance between the head vertex and the center point of the two ankle joints in the skeletal key points is obtained within a preset stable frame number after the audience enters the monitoring screen. Combined with the age classification label output by the human skeletal key point detection model, the audience's height is estimated. The ratio of the estimated audience height to the top height of the display case is calculated to obtain the height-display case matching risk coefficient.

[0145] S500: Based on the changing trends of the attitude leap index and the spatial approximation in consecutive frames, the coherence of climbing behavior is analyzed and obtained.

[0146] The step of analyzing and obtaining the coherence of climbing behavior based on the changing trends of the attitude leap index and the spatial approximation over multiple consecutive frames includes: Obtain the attitude leap index sequence and spatial approximation sequence of the audience within a preset frame window; Calculate the absolute value of the attitude rise index difference between each two adjacent frames in the attitude rise index sequence, and sum them to obtain the total change in attitude rise frame by frame. Calculate the absolute value of the spatial approximation difference between each two adjacent frames in the spatial approximation sequence, and sum them to obtain the total change in spatial approximation frame by frame; Calculate the maximum difference in attitude rise index in the attitude rise index sequence within a preset frame window as the net change in attitude rise, and calculate the maximum difference in spatial approximation in the spatial approximation sequence within a preset frame window as the net change in spatial approximation. Calculate the ratio of the net change in attitude rise to the total change in attitude rise frame by frame to obtain the attitude rise direction consistency factor, and calculate the ratio of the net change in spatial approximation to the total change in spatial approximation frame by frame to obtain the spatial approximation direction consistency factor. Calculate the arithmetic mean of the attitude jump direction consistency factor and the spatial approximation direction consistency factor to obtain the direction consistency coefficient; The product of the directional consistency coefficient and the larger of the net change in attitude elevation and the net change in spatial approximation is used as the climbing behavior coherence.

[0147] S600: Calculate the comprehensive climbing risk index based on the posture leap index, the spatial proximity, the height-display case matching risk coefficient, and the climbing behavior continuity.

[0148] The calculation of the comprehensive climbing risk index based on the posture leap index, the spatial proximity, the height-display case matching risk coefficient, and the climbing behavior continuity includes: The posture leap index, spatial proximity, height-display case matching risk coefficient, and climbing behavior coherence are weighted and calculated to obtain a comprehensive climbing risk index. The weighting factor is obtained based on the age classification label of the audience.

[0149] S700: When the comprehensive climbing risk index exceeds the preset threshold, the corresponding anti-climbing warning action is triggered.

[0150] When the comprehensive climbing risk index exceeds a preset threshold, a corresponding anti-climbing warning action is triggered, including: Based on historical audience behavior samples, a first risk threshold and a second risk threshold are obtained, wherein the first risk threshold is smaller than the second risk threshold; When the comprehensive climbing risk index is less than the first risk threshold, it is determined that the audience is within the range of normal visiting behavior, and silent monitoring is maintained. When the comprehensive climbing risk index is greater than or equal to the first risk threshold and less than the second risk threshold, a level one warning is triggered, wherein the level one warning includes at least a voice broadcast prompt; When the comprehensive climbing risk index is greater than or equal to the second risk threshold, a level-two warning is triggered, wherein the level-two warning includes at least a voice broadcast prompt and the issuance of a human intervention dispatch instruction.

[0151] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0152] This specification and accompanying drawings are merely illustrative examples of the present invention and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of the present invention. Clearly, those skilled in the art can make various alterations and modifications to the present invention without departing from its scope. If such modifications and modifications fall within the scope of the present invention and its equivalents, the present invention intends to include these modifications and modifications.

Claims

1. A machine vision-based anti-climbing early warning system for cultural relic display cases, characterized in that, include: The data acquisition module acquires the monitoring video stream of the area where the cultural relic display case is located and the physical dimension data of the cultural relic display case. The physical dimension data includes at least the height value of the top surface of the display case and the horizontal reference position of the outer side of the display case. The human skeleton key point extraction module uses a pre-trained human skeleton key point detection model to extract the coordinate sequence of human skeleton key points of the audience, wherein the human skeleton key points include at least the center point of the hip joint and the center points of both ankle joints. The posture and spatial relationship calculation module calculates the posture rise index based on the relative relationship between the vertical coordinates of the hip joint center point and the height of the top surface of the display case, and calculates the spatial approximation degree based on the distance relationship between the horizontal coordinates of the ankle joint center points and the horizontal reference position of the outer side of the display case. Specifically, it obtains the arithmetic mean of the vertical coordinates of the hip joint center point within a preset initial frame number after the viewer enters the monitoring screen, as the normal standing hip joint reference height; calculates the deviation between the hip joint height value of the current frame and the normal standing hip joint reference height, as the posture rise index; obtains the horizontal coordinates of the midpoint of the ankle joint center points in the current frame, as the horizontal position of the feet; calculates the shortest Euclidean distance between the horizontal position of the feet and the horizontal reference position of the outer side of the display case, and calculates the spatial approximation degree in combination with the display case warning distance. The height-showcase matching risk coefficient calculation module calculates the height-showcase matching risk coefficient by combining the estimated height of the audience with the height of the top surface of the showcase. The climbing behavior coherence analysis module analyzes and obtains climbing behavior coherence based on the changing trends of the posture rise index and the spatial approximation over multiple consecutive frames. Specifically, it obtains the posture rise index sequence and spatial approximation sequence of the viewer within a preset frame window; calculates the absolute value of the difference in posture rise index between each adjacent two frames in the posture rise index sequence and sums them to obtain the total change in posture rise per frame; calculates the absolute value of the difference in spatial approximation between each adjacent two frames in the spatial approximation sequence and sums them to obtain the total change in spatial approximation per frame; and calculates the maximum difference in posture rise index in the posture rise index sequence within the preset frame window as... The net change in attitude elevation is calculated, and the maximum difference in spatial approximation within the spatial approximation sequence within a preset frame window is used as the net change in spatial approximation. The ratio of the net change in attitude elevation to the total frame-by-frame change in attitude elevation is calculated to obtain the attitude elevation direction consistency factor. The ratio of the net change in spatial approximation to the total frame-by-frame change in spatial approximation is also calculated to obtain the spatial approximation direction consistency factor. The arithmetic mean of the attitude elevation direction consistency factor and the spatial approximation direction consistency factor is calculated to obtain the direction consistency coefficient. The direction consistency coefficient is multiplied by the larger value between the net change in attitude elevation and the net change in spatial approximation, and the product is used as the climbing behavior coherence. The comprehensive risk assessment module calculates a comprehensive climbing risk index based on the posture leap index, the spatial proximity, the height-display case matching risk coefficient, and the climbing behavior continuity. The early warning trigger module triggers a corresponding anti-climbing early warning action when the comprehensive climbing risk index exceeds a preset threshold.

2. The anti-climbing early warning system for cultural relic display cases based on machine vision according to claim 1, characterized in that, The acquisition of the surveillance video stream of the area where the artifact display case is located and the physical dimension data of the artifact display case includes: The monitoring equipment was used to collect surveillance video streams of the area where the artifact display case was located; Based on the display case number, the top height of the display case and the horizontal reference position of the outer side of the display case are obtained as the physical dimension data. The top height of the display case is the vertical distance of the top of the display case relative to the floor of the exhibition hall, and the horizontal reference position of the outer side of the display case is the horizontal coordinate value of the outer side of the display case facing the audience in the coordinate system of the monitoring screen.

3. The anti-climbing early warning system for cultural relic display cases based on machine vision according to claim 1, characterized in that, The pre-trained human skeletal keypoint detection model extracts the coordinate sequence of human skeletal keypoints of the audience, wherein the human skeletal keypoints include at least the center point of the hip joint and the center points of both ankle joints, including: The monitoring video stream is input frame by frame into a pre-trained human skeletal key point detection model. Human detection and skeletal key point localization are performed on each of the audience members appearing in each frame. The human skeletal key point coordinate sequence and corresponding age classification label of each audience member with the highest age confidence are output. The human skeletal key points include at least the three-dimensional spatial coordinates of the center point of the hip joint and the center points of both ankle joints. The skeletal key points of the same audience member in multiple consecutive frames are tracked and associated to form a sequence of human skeletal key point coordinates of the same audience member in the time dimension.

4. The anti-climbing early warning system for cultural relic display cases based on machine vision according to claim 3, characterized in that, The pre-training of the human skeletal key point detection model includes: Collect a dataset of sample images containing human postures of visitors in museum exhibition hall scenes, and annotate the human skeletal key points and age classification labels in each sample image to form a sample annotation set. The skeletal key points include at least the center point of the hip joint and the center points of both ankle joints of the sample. The sample image dataset is divided into child sample subsets, adult sample subsets, and elderly sample subsets according to human body shape characteristics and mobility ability; A skeletal keypoint detection model is constructed, comprising an input layer, a shared feature extraction layer, three parallel age branch layers, and an output layer. The three parallel age branch layers include a child branch, an adult branch, and an elderly branch. The skeletal keypoint detection model is jointly trained using the sample image dataset and the sample annotation set until convergence, thus obtaining the trained skeletal keypoint detection model.

5. The anti-climbing early warning system for cultural relic display cases based on machine vision according to claim 1, characterized in that, The ratio of the estimated audience height to the top height of the display case is used to calculate the height-display case matching risk coefficient, including: The stable value of the vertical distance between the head vertex and the center point of the two ankle joints in the skeletal key points is obtained within a preset stable frame number after the audience enters the monitoring screen. Combined with the age classification label output by the human skeletal key point detection model, the audience's height is estimated. The ratio of the estimated audience height to the top height of the display case is calculated to obtain the height-display case matching risk coefficient.

6. The anti-climbing early warning system for cultural relic display cases based on machine vision according to claim 1, characterized in that, The comprehensive climbing risk index is calculated based on the posture leap index, the spatial proximity, the height-display case matching risk coefficient, and the climbing behavior continuity, including: The posture leap index, spatial proximity, height-display case matching risk coefficient, and climbing behavior coherence are weighted and calculated to obtain a comprehensive climbing risk index. The weighting factor is obtained based on the age classification label of the audience.

7. The anti-climbing early warning system for cultural relic display cases based on machine vision according to claim 1, characterized in that, When the comprehensive climbing risk index exceeds a preset threshold, a corresponding anti-climbing warning action is triggered, including: Based on historical audience behavior samples, a first risk threshold and a second risk threshold are obtained, wherein the first risk threshold is smaller than the second risk threshold; When the comprehensive climbing risk index is less than the first risk threshold, it is determined that the audience is within the range of normal visiting behavior, and silent monitoring is maintained. When the comprehensive climbing risk index is greater than or equal to the first risk threshold and less than the second risk threshold, a level one warning is triggered, wherein the level one warning includes at least a voice broadcast prompt; When the comprehensive climbing risk index is greater than or equal to the second risk threshold, a level-two warning is triggered, wherein the level-two warning includes at least a voice broadcast prompt and the issuance of a human intervention dispatch instruction.

8. A machine vision-based anti-climbing early warning and identification method for cultural relic display cases, characterized in that, The anti-climbing early warning system for cultural relic display cases based on machine vision, as described in any one of claims 1-7, comprises: Acquire the surveillance video stream of the area where the cultural relic display case is located and the physical dimension data of the cultural relic display case, wherein the physical dimension data includes at least the height value of the top surface of the display case and the horizontal reference position of the outer side of the display case; A pre-trained human skeleton key point detection model is used to extract the coordinate sequence of human skeleton key points of the audience, wherein the human skeleton key points include at least the center point of the hip joint and the center points of both ankle joints. Based on the relative relationship between the vertical coordinates of the hip joint center point and the height of the top surface of the display case, a posture rise index is calculated. Furthermore, based on the distance between the horizontal coordinates of the ankle joint center points and the horizontal reference position on the outer side of the display case, a spatial approximation is calculated. Specifically: the arithmetic mean of the vertical coordinates of the hip joint center points within a preset initial frame number after the viewer enters the monitoring screen is obtained as the normal standing hip joint reference height; the deviation between the hip joint height value in the current frame and the normal standing hip joint reference height is calculated as the posture rise index; the horizontal coordinates of the midpoint of the ankle joint center points in the current frame are obtained as the horizontal position of the feet; the shortest Euclidean distance between the horizontal position of the feet and the horizontal reference position on the outer side of the display case is calculated, and combined with the display case warning distance, the spatial approximation is calculated. The height-showcase matching risk coefficient is calculated by combining the estimated audience height with the height of the top surface of the display case. Based on the changing trends of the attitude leap index and the spatial approximation in consecutive frames, the coherence of climbing behavior is analyzed and obtained. Specifically, this involves: obtaining the attitude leap index sequence and spatial approximation sequence of the viewer within a preset frame window; calculating the absolute value of the difference in attitude leap index between each adjacent two frames in the attitude leap index sequence and summing them to obtain the total change in attitude leap per frame; calculating the absolute value of the difference in spatial approximation between each adjacent two frames in the spatial approximation sequence and summing them to obtain the total change in spatial approximation per frame; and calculating the maximum difference in attitude leap index in the attitude leap index sequence within the preset frame window as the net change in attitude leap. The algorithm calculates the maximum difference in spatial approximation within a preset frame window as the net change in spatial approximation. It then calculates the ratio of the net change in attitude rise to the total frame-by-frame change in attitude rise to obtain the attitude rise direction consistency factor. Finally, it calculates the arithmetic mean of the attitude rise direction consistency factor and the spatial approximation direction consistency factor to obtain the direction consistency coefficient. The algorithm multiplies the direction consistency coefficient by the larger of the net change in attitude rise and the net change in spatial approximation, and the resulting product is used as the climbing behavior consistency degree. Based on the posture leap index, the spatial proximity, the height-display case matching risk coefficient, and the climbing behavior continuity, a comprehensive climbing risk index is calculated; when the comprehensive climbing risk index exceeds a preset threshold, a corresponding anti-climbing warning action is triggered.

Images