Smart park security early warning method and system based on Internet of Things

Abstract

The invention relates to the technical field of smart park security and protection, and discloses a smart park security and protection early warning method and system based on the Internet of Things, and the method comprises the steps: obtaining real-time meteorological monitoring data, a synchronous image sequence of a multi-angle camera, and the reading of an environment sensor; predicting the theoretical imaging characteristic range of the object surface in the current extreme environment by adopting a multi-physics field coupling algorithm and integrating atmospheric refraction, scattering and thermal distortion effects; bayesian inference is adopted to fuse the geometric deviation vector and the reflection anomaly index, and the posterior probability that the observed visual features violate the environmental adaptive physical constraint is evaluated; and dynamically adjusting the detection sensitivity according to the environmental severity by using an adaptive threshold algorithm, and when the posterior probability exceeds an environmental related threshold, generating a physical confrontation attack early warning and evidence chain. According to the scheme, the difficulty in countermeasure attack detection under the protection of an extreme environment is overcome, and reliable countermeasure detection under an all-weather condition is realized.

Description

Technical Field

[0001] This invention relates to the field of smart park security technology, and more specifically, to a smart park security early warning method and system based on the Internet of Things. Background Technology

[0002] In the field of smart park security, computer vision-based monitoring systems are widely used. However, with the development of adversarial sample technology, attackers have begun to use adversarial stickers or projection devices to launch adversarial attacks in the physical world, interfering with the normal recognition function of monitoring systems. Especially under extreme weather conditions such as heavy rain, dense fog, and strong winds, attackers will use the harsh environment as natural cover to carry out attacks.

[0003] The technical problem to be solved by this implementation is that extreme weather changes the optical propagation characteristics and the reflection characteristics of the object surface, making traditional detection methods based on physical invariance ineffective; environmental factors such as raindrops, fog, and strong light are difficult to distinguish visually from adversarial disturbances, resulting in the inability to effectively detect adversarial attacks under the cover of extreme weather. Summary of the Invention

[0004] This invention provides a smart park security early warning method and system based on the Internet of Things, which solves the technical problems of source heterogeneous data fusion, threat identification in complex scenarios, balancing early warning timeliness and accuracy, limited edge computing resources, and privacy protection requirements in related technologies.

[0005] This invention provides a smart park security early warning method based on the Internet of Things, comprising: Acquire multi-source environmental monitoring data, extract environmental disturbance parameters such as wind speed, rainfall intensity, and visibility from meteorological monitoring data, and extract texture feature maps of target areas from multi-angle image sequences; Based on the multiphysics coupling algorithm, which integrates atmospheric refraction, scattering and thermal distortion effects, the theoretical imaging feature range of the object surface under the current extreme environment is predicted, and environment-related physical constraints are generated. The actual transformation relationship of corresponding feature points in multi-angle images is calculated using a perspective geometry algorithm. The actual transformation relationship is compared with the theoretical transformation relationship considering environmental influences, and the geometric consistency deviation vector is calculated. An adaptive threshold algorithm is used to dynamically adjust the detection sensitivity according to the severity of the environment. When the norm of the geometric consistency deviation vector exceeds the environmental-related threshold, a physical adversarial attack warning signal is generated. The multiphysics coupling algorithm analyzes the impact of wind and rain conditions on the light propagation path through computational fluid dynamics simulation, calculates the imaging distortion caused by temperature gradient using the heat conduction equation, and couples the atmospheric refractive index distribution and the environmental optical transfer function to obtain a comprehensive environmental impact model.

[0006] In a preferred embodiment, the steps for acquiring multi-source environmental monitoring data include: Environmental parameters with different dimensions, such as wind speed, rainfall intensity, and visibility, are normalized and scaled to a preset range to eliminate dimensional differences. The environmental disturbance parameter vector is calculated using the time window moving average method, and a time weighting factor is introduced to give greater weight to the recent environmental state on the current detection results. The output environment perturbation parameter vector and texture feature map are used as the basis data for subsequent processing.

[0007] In a preferred embodiment, the step of generating environment-related physical constraints includes: Based on computational fluid dynamics simulation analysis of the impact of current wind and rain conditions on the light propagation path, the atmospheric refractive index distribution is calculated using the Gladstone-Dale relation, where the atmospheric refractive index is determined based on the dry air refractive constant, the water vapor refractive constant, and the air density and water vapor density at the corresponding locations. The imaging distortion caused by temperature gradient is calculated using the heat conduction equation, and the environmental optical transfer function is generated using the Kolmogorov turbulence model. By coupling atmospheric refractive index distribution and environmental optical transfer function through convolution operation, a comprehensive environmental impact model is obtained. A comprehensive distortion matrix is ​​generated by combining the thermal distortion coefficient and the light intensity attenuation factor, and the environmental adaptive physical constraints are output through the constraint condition generation function.

[0008] In a preferred embodiment, the step of calculating the geometric consistency deviation vector includes: The geometric relationship between viewpoints is calculated by using a viewpoint geometry algorithm through feature point matching and fundamental matrix estimation, thereby obtaining the actual transformation matrix of the corresponding feature points between different viewpoints. Theoretical transformation relationships are calculated based on environmental adaptive physical constraints. For each feature point in the image, it is checked whether it satisfies the constraints. Feature points that satisfy the constraints are used to calculate theoretical transformation relationships. The geometric consistency deviation vector is calculated by comparing the actual transformation relationship with the theoretical transformation relationship. Establish a geometric bias optimization objective function for the time series, minimizing the current bias while considering the continuity constraint of historical information.

[0009] In a preferred embodiment, the following steps are also included: Based on the photometric stereo algorithm, the surface reflection characteristics after environmental compensation are analyzed. The least squares method is used to solve the Lambert reflection model to identify abnormal regions that violate the Lambert reflection law and calculate the reflection anomaly index. Bayesian inference is used to fuse the geometric deviation vector and the reflection anomaly index, and the posterior probability is calculated by modeling the likelihood function through a multivariate Gaussian distribution. When the posterior probability after fusion exceeds a preset probability threshold, a physical adversarial attack warning signal is generated.

[0010] In a preferred embodiment, the step of dynamically adjusting the detection sensitivity using an adaptive threshold algorithm includes: The normalized environmental factor is calculated based on the preprocessed environmental disturbance parameter vector to ensure that each environmental parameter is a dimensionless value. The environmental severity index is calculated using a weighted summation method, with the weighting coefficients determined by optimization using the least squares method based on historical meteorological data and detection performance statistics. The adjustment factor is calculated based on the difference between the environmental severity and the reference environmental severity using a threshold adjustment function; The output dynamic detection threshold is used for subsequent anomaly determination.

[0011] In a preferred embodiment, the environmental severity index is calculated as follows: The boundary parameters are defined as follows: the upper limit of wind speed corresponding to strong winds, the upper limit of rainfall intensity corresponding to heavy rainstorms, and the lower limit of visibility corresponding to dense fog. The environmental severity index is calculated by weighting and summing the ratios of wind speed to the upper limit of wind speed, rainfall intensity to the upper limit of rainfall intensity, and the lower limit of visibility to the actual visibility. Establish a time series optimization objective function for the environmental severity index, which includes a deviation term between environmental severity and the target value and a time smoothness regularization term.

[0012] In a preferred embodiment, the step of generating a physical counter-attack warning signal includes: The coordinates of the abnormal location are determined by calculating the centroid of the deviation vector based on the spatial distribution of the geometric deviation vector. The threat level is calculated using a piecewise function based on the magnitude of the deviation vector; An environmental background description is generated based on the current environmental parameters using a template-based method. Establish a time-weighted early warning confidence optimization objective function, and output a complete structured early warning report that includes the location of the anomaly, the degree of deviation, environmental conditions, and recommended handling measures.

[0013] In a preferred embodiment, the multiphysics coupling algorithm further includes: The normalized preprocessed environmental disturbance parameter vector is input into the atmospheric transport model, and the refraction angle shift of light in different atmospheric density layers is calculated by the differential form of Snell's law. Based on the refraction angle offset and scattering coefficient, the light intensity attenuation factor of each pixel is calculated by Beer-Lambert's law, where the total optical thickness is contributed by Rayleigh scattering and Mie scattering. By combining the thermal distortion coefficient and the light intensity attenuation factor, a comprehensive distortion matrix is ​​constructed in the form of a block matrix, which includes a geometric distortion submatrix and a photometric distortion submatrix.

[0014] This invention provides a smart park security early warning system based on the Internet of Things, comprising: Meteorological monitoring stations deployed in the park are used to collect real-time meteorological monitoring data; A multi-angle camera array is used to acquire synchronized image sequences; An environmental sensor network is used to collect readings from environmental sensors. Edge computing servers are used to execute multiphysics coupling algorithms, viewpoint geometry algorithms, photometric stereo algorithms, Bayesian inference and adaptive thresholding algorithms, generate environmental adaptive physical constraints and output early warning signals.

[0015] The beneficial effects of this invention are as follows: The impact of extreme weather on the imaging process was accurately modeled through multiphysics simulation, and an environmentally adaptive physical constraint benchmark was established. After compensating for environmental interference, the non-natural characteristics of adversarial perturbations can still be identified by using multi-view geometric consistency and photometric stereo analysis. By balancing detection sensitivity and false alarm rate through Bayesian inference and adaptive threshold mechanism, the difficulty of adversarial attack detection under extreme environmental cover was overcome, and reliable adversarial detection under all-weather conditions was achieved. Attached Figure Description

[0016] Figure 1 This is a flowchart of a smart park security early warning method based on the Internet of Things according to the present invention; Figure 2 This is a time series analysis line graph of environmental monitoring data according to the present invention; Figure 3 This is a bar chart comparing the multi-camera detection performance of the present invention; Figure 4 This is a scatter plot analyzing the relationship between geometric deviation and reflection anomaly according to the present invention; Figure 5 This is a spatial distribution heatmap of anomaly detection according to the present invention; Figure 6 This is a Sankey diagram of the system early warning processing flow of the present invention. Detailed Implementation

[0017] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0018] The subject name is disclosed in at least one embodiment of the present invention, such as Figure 1 As shown, it includes the following steps: Step 1: Acquire multi-source environmental monitoring data and extract environmental disturbance parameters and target area texture features.

[0019] This step involves collecting real-time meteorological monitoring data, synchronized image sequences from multi-angle cameras, and environmental sensor readings using IoT devices. Specifically, environmental disturbance parameters such as wind speed, rainfall intensity, and visibility are extracted from the meteorological monitoring data; and texture feature maps of the target area are extracted from the multi-angle image sequences.

[0020] Data preprocessing of raw environmental monitoring data: wind speed (m / s), rainfall intensity (mm / h), visibility Environmental parameters of different dimensions (m) are normalized and scaled to the range [0,1] to eliminate the influence of dimensional differences on subsequent calculations. Output environmental disturbance parameter vector. ,in For wind speed, Rainfall intensity, For visibility; and texture feature map .

[0021] Furthermore, to reflect the impact of the time dimension on environmental monitoring, the environmental disturbance parameter vector is calculated using the time window moving average method: ,in For a moment The original environment parameter vector, The time window length is defined in seconds. This time window is set based on the timescale characteristics of environmental changes, ensuring the capture of short-term fluctuations in environmental conditions while filtering out transient noise interference. A time weighting factor is also introduced. ,in The attenuation coefficient is used as the corrected environmental parameter vector. This gives greater weight to the recent environmental conditions in the current test results; Step 2: Based on the multiphysics coupling algorithm, predict theoretical imaging features and generate environmental adaptive physical constraints.

[0022] This step comprehensively considers the impact of current environmental conditions on the imaging process. The specific processing procedure is as follows: based on environmental disturbance parameters... A multiphysics coupling algorithm is employed to integrate atmospheric refraction, scattering, and thermal distortion effects, predicting the theoretical imaging feature range of an object's surface under current extreme conditions. Environment-related physical constraints are generated. , which represents the range of imaging characteristics that a normal object should meet under the current environment.

[0023] It should be noted that the multiphysics coupling algorithm is implemented through the following sub-steps: Step 2.1: Based on computational fluid dynamics simulation analysis, analyze the impact of current wind and rain conditions on the light propagation path, and calculate the atmospheric refractive index distribution. ; Furthermore, the atmospheric refractive index distribution was calculated using the Gladstone-Dale relation: ,in m³ / kg is the refractive constant of dry air. m³ / kg is the water vapor refraction constant. For position air density at that location, Here is the density of water vapor. Air density is calculated using the ideal gas law: ,in Atmospheric pressure. J / (kg·K) is the specific gas constant. For temperature; Step 2.2: Calculate the imaging distortion caused by the temperature gradient using the heat conduction equation, and generate the environmental optical transfer function. ; Furthermore, the calculation of the environmental optical transfer function is based on atmospheric turbulence theory, using the Kolmogorov turbulence model: ,in For spatial frequency, The cutoff frequency for atmospheric turbulence is calculated using the following formula: ,in For wave number, nm is the wavelength of visible light. For transmission distance, The refractive index structure constant is calculated using the temperature gradient: ; Step 2.3: Couple the atmospheric refractive index distribution and the environmental optical transfer function to obtain a comprehensive environmental impact model.

[0024] Furthermore, the integrated environmental impact model is implemented through convolution operations: ,in For ideal imaging, This is the spatial domain form of the environmental optical transfer function. The point spread function of the refraction effect is obtained through the inverse Fourier transform: ,in This refers to the phase perturbation caused by the refractive index distribution. The aforementioned multiphysics coupling algorithm further includes the following processing steps: Step 2.4: Convert the normalized preprocessed environmental disturbance parameter vector Input the atmospheric transport model, ensure the consistency of model calculations through dimensionless parameters, and calculate the shift in the refraction angle of light at different atmospheric density layers. ; Furthermore, the refraction angle offset is calculated using the differential form of Snell's law: ,in For the first The difference between the refractive index of the atmospheric layer and the reference refractive index. The refractive index under standard atmospheric conditions. The angle of incidence is given. Atmospheric stratification employs a uniform thickness stratification method, with each layer having a thickness of [missing information]. m, atmospheric parameters within each layer are calculated using linear interpolation: ,in For the first The height of the floor; Step 2.5: Calculate the light intensity attenuation factor for each pixel based on the refraction angle offset and scattering coefficient. ; Furthermore, the light intensity attenuation factor is calculated using the Beer-Lambert law: ,in The total optical thickness is contributed by both Rayleigh scattering and Mie scattering. The Rayleigh scattering optical thickness is calculated as follows: ,in m⁻³ represents the air molecule number density. The Mie scattering optical thickness is related to the aerosol concentration. ,in For extinction cross section, For the Angstrom index, The aerosol number concentration is obtained by inverting visibility data: ; Step 2.6: Combine the thermal distortion coefficient and the light intensity attenuation factor to generate a comprehensive distortion matrix. Output environment adaptive physical constraints ,in The constraint generation function is specifically defined as follows: ,in This is a vector of deviation thresholds set based on environmental conditions.

[0025] Furthermore, the integrated distortion matrix The construction adopts a block matrix form: ,in For geometrically distorted submatrices, This is the photometric distortion submatrix. The geometric distortion submatrix is ​​constructed using the refraction angle offset: ,in For the Kronecker function, These are the geometric distortion weighting coefficients. The photometric distortion submatrix is ​​constructed using the light intensity attenuation factor: ,in This is the photometric distortion weighting coefficient.

[0026] Furthermore, to reflect the impact of temporal evolution on the distortion matrix, a time-dependent distortion matrix update mechanism is established: ,in For update rate parameter, Based on current environmental parameters Calculate the new distortion matrix. The values ​​of the matrix elements are constrained as follows: (Geometric distortion coefficient range constraint) (Optical distortion coefficient range constraint) (Matrix nonsingularity constraint) (Time-varying smoothness constraint), where It is the Frobenius norm; Furthermore, the deviation threshold vector The calculation method is as follows: ,in This is the baseline threshold vector under standard conditions, with each component taking values ​​in the range of [0.001, 0.1]. This is the environmental sensitivity coefficient, with a value range of [0.1, 2.0]. The environmental severity index ranges from [0, 1]. This calculation method ensures that constraints are appropriately relaxed under harsh environmental conditions and tightened under favorable environmental conditions, thereby achieving dynamic threshold adjustment that adapts to the environment.

[0027] Step 3: Calculate the multi-view geometric consistency deviation and detect anomalies that violate environmental constraints.

[0028] This step utilizes a perspective geometry algorithm to calculate the actual transformation relationships of corresponding feature points in multi-angle images. The actual transformation relationship is compared with the theoretical transformation relationship that considers environmental impact. Perform a comparison and calculate the geometric consistency deviation vector: in Based on environmental adaptive physical constraints The calculated theoretical transformation relationships. Environmental constraints. The application process is as follows: for each feature point in the image Check whether it meets the constraints. Feature points that satisfy the constraints are considered to conform to normal imaging characteristics under the current environment and are used to calculate theoretical transformation relationships. .

[0029] Furthermore, to reflect the impact of the time dimension on geometric consistency analysis, an objective function for optimizing the geometric deviation of the time series is established: The first term is the time-weighted cumulative deviation term. As time weight, The first term is the time decay parameter; the second term is the time continuity regularization term. For regularization parameters, Let be the length of the time series. The objective function minimizes the current deviation while considering the continuity constraint of historical information. The constraint conditions are: (Non-negativity constraint); (Upper bound constraint); (Cumulative constraints); The aforementioned perspective geometry algorithm takes a multi-view image sequence as input. and camera calibration parameter matrix The output is the actual transformation matrix of the corresponding feature point between different viewpoints. The algorithm calculates the geometric relationship between viewpoints through feature point matching and fundamental matrix estimation.

[0030] Furthermore, feature point matching employs the SIFT feature descriptor and nearest neighbor distance ratio method: first, SIFT feature points are extracted from images at each viewpoint to obtain feature point coordinates. and 128-dimensional descriptor vector Then calculate the Euclidean distance between the descriptors. The Lowe's ratio test is used for matching and filtering: when When a match is considered valid, and These are the nearest neighbor and second nearest neighbor distances, respectively. The fundamental matrix is ​​estimated using the RANSAC algorithm: 8 pairs of matching points are randomly selected, and the fundamental matrix is ​​calculated using an 8-point algorithm. Using epipolar constraints Verify interior points, where and These are homogeneous coordinates. The actual transformation matrix is ​​obtained through essential matrix decomposition: Then perform SVD decomposition. Rotation matrix Translation vector pass and Calculation, where ; In this embodiment of the application, to improve the robustness of detection, the following steps are also included: Step 3.1: Analyze the surface reflection characteristics after environmental compensation based on the photometric stereo algorithm, identify anomalous areas that still violate Lambert's law of reflection under the current lighting and weather conditions, and calculate the reflection anomaly index. ; Step 3.2: Use Bayesian inference to fuse the geometric bias vector and reflection anomaly index Calculate the posterior probability .

[0031] The aforementioned photometric stereo algorithm takes as input a set of images under multi-light source conditions. and the direction vector of the light source The output is the surface normal field. and reflection anomaly index The photometric stereo algorithm reconstructs surface geometry by analyzing pixel intensity variations under different lighting conditions.

[0032] Furthermore, the photometric stereo algorithm employs the least squares method to solve the Lambertian reflection model: for each pixel position Establish a system of linear equations , in These are pixel intensity vectors under different lighting conditions. for The light source direction matrix, Let be the surface normal vector to be determined. Solve using least squares. Obtain the normal vector, then normalize it: The reflection anomaly index is calculated by fitting the residuals: The reflection anomaly index quantifies the degree of deviation between actual reflection characteristics and the Lambertian reflection model. It was initially believed that there was an abnormality in the reflection. Furthermore, to reflect the impact of the time dimension on reflection anomaly detection, a time-series reflection anomaly optimization objective function is established: in As a time-weighted factor, For attenuation parameters, The baseline anomaly index for normal reflection is... This represents the length of the time window for reflection analysis. This is a time continuity parameter. Constraints include: (Non-negativity constraint) (Upper bound constraint) (Cumulative anomaly constraints) (Rate of change constraint) ensures the physical rationality and temporal continuity of the reflection anomaly index; The input to the aforementioned Bayesian inference is the geometric bias vector. and reflection anomaly index and prior probability The output is the posterior probability. .

[0033] Furthermore, the likelihood function Modeling using a multivariate Gaussian distribution: ,in To observe the feature vector, The feature dimension, This is the mean vector under attack conditions. The parameters are obtained through maximum likelihood estimation of historical attack samples, representing the covariance matrix. The likelihood function under normal conditions is also shown. Using the same form, but with the mean vector as The covariance matrix is ; Furthermore, to reflect the impact of the time dimension on Bayesian inference, a posterior probability optimization objective function for time series is established: in As time weight, The attenuation coefficient is... The time window length for Bayesian inference. Constraints include: (Probability value constraints); (Probability normalization constraint); (Smoothness constraint of probability change); (Average attack probability upper limit constraint) ensures the temporal continuity and statistical rationality of posterior probability; Furthermore, prior probability Based on the statistics of historical security incidents in the park, the value is 0.05, which represents the baseline probability of a counterattack occurring under normal operating conditions. The prior probability is calculated by analyzing the security log data of the park over the past 12 months. Step 4: Dynamically adjust the detection threshold based on the severity of the environment to generate an early warning of adversarial attacks.

[0034] This step utilizes an adaptive threshold algorithm to dynamically adjust the detection sensitivity based on environmental severity. Specifically, it is based on environmental disturbance parameters. Calculate the environmental severity index Dynamically adjust detection threshold ,in As the baseline threshold, This is the threshold adjustment function.

[0035] The aforementioned adaptive thresholding algorithm includes the following processing steps: Step 4.1: Based on the preprocessed environmental disturbance parameter vector Calculate normalized environmental factors Ensure that all environmental parameters are dimensionless values; Furthermore, the boundary parameters are defined as follows: m / s (the maximum wind speed corresponding to a level 10 gale). mm / h (corresponding to the upper limit of heavy rainstorm intensity) m (corresponding to the lower limit of visibility in dense fog), these boundary parameters are determined based on meteorological standards to ensure that the normalization process can cover various extreme weather conditions that the park may encounter; Step 4.2: Calculate the environmental severity index using weighted summation; ,in The weighting coefficients are satisfied. The normalization process eliminates the influence of inconsistent dimensions of different environmental parameters on the weighted summation operation; Furthermore, the weighting coefficients are determined by optimizing them using the least squares method based on historical meteorological data and detection performance statistics. (Weight of wind speed influence) (Weight of rainfall impact) (Visibility influence weights), all weight coefficients are positive and satisfy normalization constraints. The allocation of weighting coefficients reflects the relative impact of different environmental factors on the detection performance of the visual monitoring system; Furthermore, the time series optimization objective function for the environmental severity index is: The first item is the deviation between the environmental severity and the target value. The target severity is calculated directly based on environmental parameters. The first term is the length of the historical time window; the second term is the time smoothing regularization term. These are the smoothness weight parameters. Constraints include: (Value range constraints); (Rate of change constraint); (Minimum average severity constraint) ensures the temporal continuity and physical rationality of the environmental severity index; Step 4.3: Adjusting the threshold function Calculate the adjustment factor, where To adjust the sensitivity parameters, For reference to environmental severity; Furthermore, the parameter values ​​are defined as follows: The value range is [0.5, 1.5]. Adjusting the sensitivity parameter controls the degree of adjustment of the detection threshold due to environmental changes. , represents the environmental severity baseline value under normal operating conditions of the park, with a value range of [0.2, 0.5]. The environmental severity baseline value is determined by statistically analyzing the 75th percentile of the park's historical environmental data. Step 4.4: Output dynamic detection threshold This is used for subsequent anomaly detection.

[0036] Furthermore, the baseline threshold The value range is [0.1, 0.3]. The benchmark threshold is determined based on the requirement that the false alarm rate under standard environmental conditions be controlled within 5%, so as to ensure that the system has high detection accuracy under normal environmental conditions. Furthermore, to reflect the impact of the time dimension on dynamic threshold adjustment, a time-dependent threshold optimization objective function is established: in As time weight, For attenuation parameters, The optimal threshold is calculated based on the current environmental severity. The time window for threshold adjustment This is the time smoothing parameter. Constraints include: (Threshold value range constraints); (Threshold change rate constraint); (Minimum average threshold constraint); (Environment-related upper bound constraints) ensure the rationality and stability of dynamic threshold adjustment; When the norm of the geometric consistency deviation vector Exceeding environmentally relevant thresholds When, or when the posterior probability after fusion When the probability exceeds a preset threshold, a physical attack warning signal is generated.

[0037] Furthermore, the preset probability threshold is defined as follows: The value range is [0.6, 0.9]. The preset probability threshold is determined based on the principle of balancing detection sensitivity and false alarm rate. When the posterior probability exceeds the preset probability threshold, it indicates that there is a high confidence in the existence of adversarial attacks, and an early warning mechanism needs to be triggered immediately. The process of generating and decoding early warning signals includes: converting mathematical calculation results into executable early warning information, specifically, based on the geometric deviation vector. The spatial distribution determines the coordinates of the anomaly location, and the threat level (low / medium / high) is calculated based on the magnitude of the deviation vector, according to the current environmental parameters. Generate an environmental background description and finally output a structured early warning report, which includes a complete chain of evidence including the location of the anomaly, the degree of deviation, environmental conditions, and recommended remedial measures.

[0038] Furthermore, the coordinates of the abnormal location are calculated using the centroid of the deviation vector: ,in For pixel coordinates, This represents the deviation value at that location. Threat level determination uses a piecewise function: when... At that time, it was a low-level threat. At that time, it was a medium-level threat. This is a high-level threat. The environmental background description is generated using a template-based method: based on... Select the corresponding environment description template when the value is selected. When described as a "good environment", It was described as a "moderately harsh environment" when The description is "severe conditions," with specific wind speed, rainfall, and visibility values ​​attached. Furthermore, to reflect the impact of the time dimension on the generation of early warning signals, a time-weighted early warning confidence optimization objective function is established: ,in For a moment The confidence level of the early warning As a time-weighted factor, For time decay parameters, This is a detection indicator function (1 if an anomaly is detected, 0 otherwise). The length of the time window for generating the early warning. Constraints include: (when or Minimum confidence constraint at time). (Confidence level change rate constraint); (Maximum number of warnings within the time window) (Confidence upper bound constraint based on deviation ratio, where) This ensures the rationality and operability of the early warning signal generation; This case study uses a security monitoring scenario in a technology park during a nighttime downpour. From 8:30 PM to 9:00 PM on June 15, 2024, the park experienced severe convective weather, with wind speeds reaching 12 m / s, rainfall intensity of 18 mm / h, and visibility dropping to 80 m. Under the cover of this adverse environment, an intruder attempted to interfere with the facial recognition system at the park's east entrance using anti-scratching stickers.

[0039] Step 1 processing results: Environmental disturbance parameter extraction; Taking the data at 20:40:00 as an example, the original environmental parameters are: m / s, mm / h, m.

[0040] The environmental disturbance parameter vector is obtained after normalization. Time-weighted environmental parameter vector: ; Texture feature maps extracted from multi-angle image sequences Texture response maps with 8 directions and 3 scales were obtained by extracting using Gabor filter banks, with dimensions of 1024×768×24.

[0041] Figure 2 The dynamic changes of environmental disturbance parameters during the period from 20:30 to 20:50 are shown, including three key environmental factors: wind speed, rainfall intensity, and visibility.

[0042] Step 2 processing result: Generation of environmental adaptive physical constraints; Based on environmental disturbance parameters Calculation results of multiphysics coupling algorithm; Atmospheric refractive index distribution: Ambient optical transfer function cutoff frequency: m⁻¹ Comprehensive distortion matrix After the construction is complete, environmental adaptive physical constraints are generated: ; This constraint defines the range of imaging features that normal feature points should satisfy under the current heavy rain environment.

[0043] Step 3 processing results: Multi-view geometric consistency deviation detection; First, apply environmental constraints. Feature point selection: Of the 1594 detected feature points, 1247 met the constraints. These feature points are used to calculate theoretical transformation relationships. .

[0044] The actual transformation matrix is ​​calculated through SIFT feature matching and fundamental matrix estimation. With theoretical transformation matrix Deviation: Geometric consistency deviation vector calculation: ; Reflectance anomaly index detected by photometric stereo algorithm: ; Bayesian inference calculates posterior probabilities: Observation feature vector ; Likelihood function calculation: , ; Marginal probability: Posterior probability: ; Figure 3 The comparison of detection performance metrics for four cameras at different angles is presented, including the number of detected feature points, the number of effective matching point pairs, the average confidence score, and the image sharpness score.

[0045] Figure 4 The distribution relationship between the geometric consistency deviation vector and the reflection anomaly index is shown, and the effectiveness of the Bayesian inference fusion algorithm is verified by the scatter distribution.

[0046] Step 4 processing results: Dynamic threshold adjustment and early warning generation; Due to geometric deviation (Meets the warning conditions), but the posterior probability (Probability threshold not met), triggering an early warning based on the geometric consistency deviation judgment system.

[0047] In this example, the system detected a suspected adversarial attack anomaly at 20:40:15 based on geometric consistency deviation, with a detection delay of 0.8 seconds. Although the Bayesian posterior probability was low (0.154), the geometric deviation significantly exceeded the dynamic threshold (0.368 vs 0.168), indicating the presence of a geometric inconsistency anomaly. Through environmental adaptive constraint filtering, the system effectively distinguished between environmental interference and human-induced adversarial attacks in a rainstorm environment.

[0048] Figure 5 The anomaly detection spatial distribution in the eastern entrance area of ​​the park is shown, with the color intensity representing the strength of the detected anomalies.

[0049] Figure 6 The patented system demonstrates the complete data flow and processing workflow from environmental monitoring to final early warning output, including key steps such as multi-source data acquisition, algorithm processing, anomaly detection, and early warning generation.

[0050] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A smart park security early warning method based on the Internet of Things, characterized in that, Includes the following steps: Acquire multi-source environmental monitoring data, extract environmental disturbance parameters such as wind speed, rainfall intensity, and visibility from meteorological monitoring data, and extract texture feature maps of target areas from multi-angle image sequences; Based on the multiphysics coupling algorithm, which integrates atmospheric refraction, scattering and thermal distortion effects, the theoretical imaging feature range of the object surface under the current extreme environment is predicted, and environment-related physical constraints are generated. The actual transformation relationship of corresponding feature points in multi-angle images is calculated using a perspective geometry algorithm. The actual transformation relationship is compared with the theoretical transformation relationship considering environmental influences, and the geometric consistency deviation vector is calculated. An adaptive threshold algorithm is used to dynamically adjust the detection sensitivity according to the severity of the environment. When the norm of the geometric consistency deviation vector exceeds the environmental-related threshold, a physical adversarial attack warning signal is generated. The multiphysics coupling algorithm analyzes the impact of wind and rain conditions on the light propagation path through computational fluid dynamics simulation, calculates the imaging distortion caused by temperature gradient using the heat conduction equation, and couples the atmospheric refractive index distribution and the environmental optical transfer function to obtain a comprehensive environmental impact model.

2. The smart park security early warning method based on the Internet of Things according to claim 1, characterized in that, The steps for acquiring multi-source environmental monitoring data include: Environmental parameters with different dimensions, such as wind speed, rainfall intensity, and visibility, are normalized and scaled to a preset range to eliminate dimensional differences. The environmental disturbance parameter vector is calculated using the time window moving average method, and a time weighting factor is introduced to give greater weight to the recent environmental state on the current detection results. The output environment perturbation parameter vector and texture feature map are used as the basis data for subsequent processing.

3. The smart park security early warning method based on the Internet of Things according to claim 1, characterized in that, The steps for generating the physical constraints related to the environment include: Based on computational fluid dynamics simulation analysis of the impact of current wind and rain conditions on the light propagation path, the atmospheric refractive index distribution is calculated using the Gladstone-Dale relation, where the atmospheric refractive index is determined based on the dry air refractive constant, the water vapor refractive constant, and the air density and water vapor density at the corresponding locations. The imaging distortion caused by temperature gradient is calculated using the heat conduction equation, and the environmental optical transfer function is generated using the Kolmogorov turbulence model. By coupling atmospheric refractive index distribution and environmental optical transfer function through convolution operation, a comprehensive environmental impact model is obtained. A comprehensive distortion matrix is ​​generated by combining the thermal distortion coefficient and the light intensity attenuation factor, and the environmental adaptive physical constraints are output through the constraint condition generation function.

4. The smart park security early warning method based on the Internet of Things according to claim 1, characterized in that, The steps for calculating the geometric consistency deviation vector include: The geometric relationship between viewpoints is calculated by using a viewpoint geometry algorithm through feature point matching and fundamental matrix estimation, thereby obtaining the actual transformation matrix of the corresponding feature points between different viewpoints. Theoretical transformation relationships are calculated based on environmental adaptive physical constraints. For each feature point in the image, it is checked whether it satisfies the constraints. Feature points that satisfy the constraints are used to calculate theoretical transformation relationships. The geometric consistency deviation vector is calculated by comparing the actual transformation relationship with the theoretical transformation relationship. Establish a geometric bias optimization objective function for the time series, minimizing the current bias while considering the continuity constraint of historical information.

5. The smart park security early warning method based on the Internet of Things according to claim 1, characterized in that, It also includes the following steps: Based on the photometric stereo algorithm, the surface reflection characteristics after environmental compensation are analyzed. The least squares method is used to solve the Lambert reflection model to identify abnormal regions that violate the Lambert reflection law and calculate the reflection anomaly index. Bayesian inference is used to fuse the geometric deviation vector and the reflection anomaly index, and the posterior probability is calculated by modeling the likelihood function through a multivariate Gaussian distribution. When the posterior probability after fusion exceeds a preset probability threshold, a physical adversarial attack warning signal is generated.

6. The smart park security early warning method based on the Internet of Things according to claim 1, characterized in that, The step of dynamically adjusting the detection sensitivity using the adaptive threshold algorithm includes: The normalized environmental factor is calculated based on the preprocessed environmental disturbance parameter vector to ensure that each environmental parameter is a dimensionless value. The environmental severity index is calculated using a weighted summation method, with the weighting coefficients determined by optimization using the least squares method based on historical meteorological data and detection performance statistics. The adjustment factor is calculated based on the difference between the environmental severity and the reference environmental severity using a threshold adjustment function; The output dynamic detection threshold is used for subsequent anomaly determination.

7. The smart park security early warning method based on the Internet of Things according to claim 6, characterized in that, The method for calculating the environmental severity index is as follows: The boundary parameters are defined as follows: the upper limit of wind speed corresponding to strong winds, the upper limit of rainfall intensity corresponding to heavy rainstorms, and the lower limit of visibility corresponding to dense fog. The environmental severity index is calculated by weighting and summing the ratios of wind speed to the upper limit of wind speed, rainfall intensity to the upper limit of rainfall intensity, and the lower limit of visibility to the actual visibility. Establish a time series optimization objective function for the environmental severity index, which includes a deviation term between environmental severity and the target value and a time smoothness regularization term.

8. The smart park security early warning method based on the Internet of Things according to claim 1, characterized in that, The steps for generating a physical counter-attack warning signal include: The coordinates of the abnormal location are determined by calculating the centroid of the deviation vector based on the spatial distribution of the geometric deviation vector. The threat level is calculated using a piecewise function based on the magnitude of the deviation vector; An environmental background description is generated based on the current environmental parameters using a template-based method. Establish a time-weighted early warning confidence optimization objective function, and output a complete structured early warning report that includes the location of the anomaly, the degree of deviation, environmental conditions, and recommended handling measures.

9. The smart park security early warning method based on the Internet of Things according to any one of claims 1 to 8, characterized in that, The multiphysics coupling algorithm also includes: The normalized preprocessed environmental disturbance parameter vector is input into the atmospheric transport model, and the refraction angle shift of light in different atmospheric density layers is calculated by the differential form of Snell's law. Based on the refraction angle offset and scattering coefficient, the light intensity attenuation factor of each pixel is calculated by Beer-Lambert's law, where the total optical thickness is contributed by Rayleigh scattering and Mie scattering. By combining the thermal distortion coefficient and the light intensity attenuation factor, a comprehensive distortion matrix is ​​constructed in the form of a block matrix, which includes a geometric distortion submatrix and a photometric distortion submatrix.

10. A smart park security early warning system based on the Internet of Things (IoT), used to execute the smart park security early warning method based on the IoT as described in any one of claims 1-9, characterized in that, include: Meteorological monitoring stations deployed in the park are used to collect real-time meteorological monitoring data; A multi-angle camera array is used to acquire synchronized image sequences; An environmental sensor network is used to collect readings from environmental sensors. Edge computing servers are used to execute multiphysics coupling algorithms, viewpoint geometry algorithms, photometric stereo algorithms, Bayesian inference and adaptive thresholding algorithms, generate environmental adaptive physical constraints and output early warning signals.

Images