Frontier tourism safety risk ai early warning method

Abstract

The application discloses a border tourism safety risk early warning method and relates to the technical field of tourism safety monitoring. The method specifically comprises the following steps: generating a risk field that changes with real-time environmental data for the border of a border control area without physical isolation in a border tourism scenic area, wherein the risk field gives a continuous risk value to each position in the area; collecting position data of tourists and calculating instantaneous speed and acceleration to obtain the risk gradient of the current position; establishing a force balance model to express the instantaneous acceleration as the resultant force of the internal driving force, the risk gradient counterforce and the motion damping force, and solving the internal driving force; orthogonally decomposing the internal driving force to obtain the risk direction component force, performing statistical filtering on the risk direction component force, taking the filtered value as the driving force to perform forward trajectory deduction, obtaining the probability of entering a high-risk area and the expected arrival time, and outputting a graded early warning. Through mechanical analysis of the motion driving force of the tourists, the application can realize active early warning before the border crossing behavior occurs.

Description

Technical Field

[0001] This invention relates to the field of tourism safety monitoring technology, and more specifically, to an AI-based early warning method for tourism safety risks in border areas. Background Technology

[0002] Border control zones in border tourist areas are often soft borders that are legally off-limits but lack physical barriers, such as border buffer zones or the perimeter of military restricted areas. These areas lack physical barriers like fences and warning signs, making it difficult for tourists to visually identify the boundary location. During their sightseeing, they may unintentionally approach or even mistakenly enter high-risk areas.

[0003] Currently, there are two main methods for monitoring and early warning in such scenarios. One is a tourist safety early warning method based on location information (publication number CN116347354A). This method collects the real-time location of tourists to determine whether they are within a safe distance and the duration of their stay. If they exceed the safe distance or stay beyond the designated time, an early warning message is sent to the tourist or relevant personnel. This method uses the spatial relationship between the location and a preset safety boundary as the basis for early warning. When the tourist's location coordinates indicate that they have exceeded the safe range, an alarm is triggered. By the time the alarm is triggered, the boundary crossing has already occurred.

[0004] Another approach is a multi-feature trajectory prediction and cross-border defense early warning method for border and coastal defense control targets (publication number CN121011112B). This method extracts multi-dimensional trajectory features of the controlled targets, constructs a trajectory prediction model, and outputs subsequent trajectories under preset constraints, thereby predicting the future trajectory of the controlled targets and identifying cross-border defense risks. This method relies on a large number of labeled samples for model training, but in border tourism scenarios, cross-border samples are scarce, making it difficult to obtain sufficient samples to support model training. At the same time, tourists' normal sightseeing behavior near the border and their approaching behavior before crossing the border are highly similar in terms of motion appearance, making it difficult to effectively distinguish between the two by relying on trajectory pattern classification.

[0005] Neither of the above two methods can provide accurate risk warnings before tourists cross the border; therefore, an AI-based early warning method for border tourism safety risks is proposed to address these issues. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present invention provide an AI early warning method for border tourism safety risks. The technical problem to be solved is that the existing early warning methods have problems such as post-event response or difficulty in distinguishing between normal sightseeing and border crossing behavior in the soft boundary scenario of border tourist attractions, and it is difficult to accurately identify risks before tourists cross the boundary.

[0007] To achieve the above objectives, the present invention provides a method for early warning of safety risks in border tourism, comprising the following steps.

[0008] S1: For border control zones without physical barriers in border tourist areas, a risk field is generated that changes with real-time environmental data. This risk field assigns a continuous risk value to each location within the area. The continuously varying risk values ​​describe the change in risk level from the boundary inland, with the risk value of a tourist's location gradually increasing as they approach the boundary.

[0009] Furthermore, the risk field is generated by using the border control zone boundary as the source and employing an anisotropic diffusion method to generate the basic risk field.

[0010] The diffusion coefficient perpendicular to the boundary is greater than that parallel to the boundary, creating a significant gradient in risk value perpendicular to the boundary, while the change is gradual parallel to the boundary. This distinguishes two behavioral patterns in terms of spatial risk distribution: those who walk directly towards the boundary and those who sightsee along the boundary. The risk value increases rapidly at the former's location, while it remains relatively constant at the latter.

[0011] The basic risk field is modulated by using a preset mapping model to output amplitude and bandwidth parameters based on the input real-time environmental factors, thereby obtaining the risk field and dynamically adjusting the amplitude and spatial range of the risk field according to the real-time environment.

[0012] The real-time environmental factors include visibility data and border control level data. When visibility decreases or the border control level increases, the bandwidth parameter increases, and the high-risk area expands to both sides of the border, thus increasing the advance warning time accordingly.

[0013] S2: Real-time acquisition of tourist location data, calculation of tourist instantaneous velocity and instantaneous acceleration, and acquisition of the risk gradient of the tourist's current location within the risk field. The risk gradient points in the direction of the fastest increase in risk value, reflecting the risk change trend of the tourist's location.

[0014] S3: Establish a force balance model, representing the instantaneous acceleration as the resultant force of the internal driving force, the reaction force of the risk gradient, and the motion damping force. Substitute the known instantaneous acceleration, the risk gradient, and the motion damping force to solve for the internal driving force.

[0015] This model treats the risk field as an external force field that exerts a repulsive force on tourists. By using force balance relationships, it transforms the internal driving force, which cannot be directly observed, into a solvable mechanical quantity, making it possible to analyze the motivation for movement from the appearance of motion.

[0016] Furthermore, the motion damping force is obtained by multiplying the damping coefficient by the instantaneous velocity, and the value of the damping coefficient is related to the magnitude of the instantaneous velocity. When the instantaneous velocity is lower than a preset low-speed threshold, the value of the damping coefficient increases, thereby suppressing the interference of positioning noise on the driving force solution when the tourist is in a low-speed or hesitant state.

[0017] After solving for the internal driving force, the covariance matrix of the internal driving force is calculated based on the positioning error propagation to quantify the uncertainty of the driving force solution and provide noise information for subsequent filtering processing.

[0018] S4: Using the direction of the risk gradient as the risk direction and the direction perpendicular to the risk gradient as the safety direction, orthogonally decompose the internal driving force to obtain the risk direction component and the safety direction component. This distinguishes the component of the tourist driving force pointing towards the high-risk area from the component along the boundary safety direction.

[0019] When tourists are engaged in normal sightseeing activities, the risk directional force is small or negative.

[0020] When tourists intend to approach restricted areas, the risk directional force increases positively.

[0021] Even when the trajectory of motion has not yet shown obvious signs of crossing the boundary, risk tendency can be identified from the composition of the driving force.

[0022] S5: Perform statistical filtering on the time series of the risk direction component to obtain the filtered risk direction component and its estimated variance, and extract a stable intention trend from the observation sequence affected by noise. The observation noise variance of the statistical filtering is taken as the projection of the covariance matrix on the risk direction, so that the filtering intensity is adaptively adjusted with the positioning error.

[0023] The statistical filtering is implemented using a Kalman filter, with the risk directional force as the observation value and the random walk process as the state equation. The posterior estimate is output through a two-step recursive process of prediction and update.

[0024] Following S5, the time series of the filtered risk direction component is accumulated and detected. When a statistically significant positive jump is detected in the filtered risk direction component, an intent change marker is generated, thereby capturing a sudden increase in the tourist's risk intent.

[0025] Simultaneously, when the estimated variance consistently exceeds a preset first threshold, and the risk value of the tourist's current location exceeds a preset second threshold, the tourist is determined to be in an ambiguous intent state, and a attention prompt is generated. An ambiguous intent state corresponds to the tourist's behavior of repeatedly lingering near the boundary and frequently changing direction; recognizing this state helps detect tentative behaviors before crossing the boundary.

[0026] S6: Using the filtered risk direction component as the driving force, and combining it with the risk field, a forward trajectory is extrapolated to obtain the probability of tourists entering high-risk areas and their estimated arrival time. This allows for the prediction of future risks without relying on the actual trajectory.

[0027] Furthermore, the forward trajectory extrapolation method is as follows: taking the tourist's current position and instantaneous speed as the initial state, and using the filtered risk direction component and the estimated variance as driving force parameters, multiple virtual trajectories are generated in the risk field.

[0028] The impact of driving force uncertainty on the trajectory is simulated by sampling from the distribution characterized by the estimated variance. The proportion of the virtual trajectory that touches a preset high-risk contour surface is statistically analyzed as the probability, and the average time of contact is calculated as the expected arrival time.

[0029] S7: Output a graded warning based on the probability and the expected arrival time.

[0030] Furthermore,

[0031] When the probability is lower than a preset lower limit and the filtered risk direction component is less than or equal to zero, it is determined to be safe and no warning is triggered.

[0032] When the probability exceeds a preset upper limit and the expected arrival time is greater than a preset safety intervention time, a yellow warning is issued.

[0033] When the probability exceeds the preset upper limit and the estimated arrival time is less than or equal to the safety intervention time, a red alert is output, and directional guidance is given to intervene in the opposite direction of the risk gradient, where the opposite direction is the direction in which the risk value decreases the fastest.

[0034] When the intent change marker is generated and the probability exceeds a preset threshold, a red alert and directional guidance are also output.

[0035] Early warning levels are determined by both probability and time factors, taking into account both the timeliness of the warning and the feasibility of intervention.

[0036] This application constructs a risk field and force balance model, inversely solves for the internal driving force from tourist movement data, and performs orthogonal decomposition on it, distinguishing between normal sightseeing and intentional border crossing without relying on samples of border crossing behavior. By extrapolating forward trajectory, it obtains the risk probability and estimated arrival time before border crossing behavior occurs, achieving proactive early warning of border crossing risks in soft boundary scenarios of border tourism.

[0037] The technical effects and advantages of this invention are as follows:

[0038] This invention constructs a continuously changing risk field to quantify the risk distribution in scenic areas numerically, and uses a force balance model to inversely solve for the internal driving force from tourist location data. This model incorporates both the repulsive effect of the risk field on tourists and motion damping, separating the driving force generated by the tourists themselves, and then orthogonally decomposing it to obtain the risk direction component. Statistical filtering of the risk direction component is applied to suppress location noise interference, extract stable intention trends, and then the filtered component is used as the driving force for forward trajectory extrapolation. During extrapolation, multiple virtual trajectories are generated within the risk field, the proportion of trajectories touching high-risk areas is statistically analyzed to obtain the probability, and the average of the first contact time is calculated to obtain the estimated arrival time, thus completing risk prediction before the actual boundary crossing behavior occurs. This method does not rely on boundary crossing behavior samples for pattern learning; it directly assesses the tourist's movement intention through mechanical analysis, which can improve the problem of insufficient timely warnings in soft boundary scenarios.

[0039] This invention orthogonally decomposes the internal driving force, dividing the total driving force into risk-direction and safety-direction components. When tourists are engaged in normal sightseeing activities near the boundary, their risk-direction component is small or negative, indicating no driving force pointing towards high-risk areas; when tourists intend to approach the restricted area, the risk-direction component increases positively. Even if the movement trajectory does not show obvious boundary-crossing characteristics, risk tendency can be identified from the composition of the driving force. Simultaneously, continuous monitoring of the filter variance combined with the judgment of location risk values ​​can identify tentative lingering behavior of tourists near the boundary and generate attention prompts. This mechanism does not rely on pre-learning of boundary-crossing behavior patterns, thus solving the problem of false alarms caused by the similarity between normal sightseeing activities and pre-boundary-crossing behavior in movement appearance. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the overall process of implementing the method of the present invention;

[0041] Figure 2 This is the process for detecting sudden changes and determining ambiguous states in accordance with the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Example 1

[0044] As attached Figures 1 to 2The AI-based early warning method for border tourism safety risks, as shown, abstracts the scenic area into a continuously changing risk field. It considers the movement of tourists as the result of a mechanical response under the combined action of their internal driving force and the risk field. The method inversely solves for this internal driving force from observable kinematic data, and then, based on the direction, intensity, and time-varying pattern of the driving force, achieves tiered early warning through forward trajectory deduction. The following provides a detailed explanation of each step of this method.

[0045] S1: Generate the risk field

[0046] S1 generates a risk field that changes with real-time environmental data for border control zones without physical barriers in border tourist areas. The risk field assigns a continuous risk value to each location within the area; the higher the risk value, the greater the risk of crossing the boundary.

[0047] First, geometric data of the border control zone boundary is acquired, which describes the spatial orientation of the boundary line as an ordered sequence of coordinate points. Using this boundary line as the source, an anisotropic diffusion method is employed to generate a basic risk field.

[0048] The scenic area is discretized into a uniform grid. The grid spacing is determined based on the positioning accuracy and the area of ​​the scenic area. The grid spacing is taken as... Set the risk value for all grid points on the boundary line to the highest value. , Take 100.

[0049] Anisotropic diffusion equations in the inland direction Perform iterative solution, where This is the risk value. For iteration time, For the diffusion tensor.

[0050] The diffusion tensor takes different diffusion coefficients in the direction perpendicular to the boundary and the direction parallel to the boundary, and the diffusion coefficient along the direction perpendicular to the boundary is... Greater than the diffusion coefficient along the direction parallel to the boundary. The iteration stops when the maximum change in risk value between two consecutive iterations is less than the preset convergence threshold, or when the preset maximum number of iterations is reached.

[0051] After iteration to steady-state convergence, the resulting basic risk field exhibits a significant risk value gradient in the direction perpendicular to the boundary, while its change is gradual in the direction parallel to the boundary. This method can accurately characterize the spatial risk distribution under complex boundary shapes and is suitable for scenarios with tortuous boundaries requiring high-precision risk modeling.

[0052] As another implementation method, when the boundary lines are relatively regular and computational efficiency is critical, the basic risk field can also be constructed using an analytical distance function. For any location in space, the shortest distance from that location to each line segment of the border control zone boundary is calculated. Substituting into the exponential decay function:

[0053] The risk value is obtained, where The attenuation characteristic length. The settings are based on the actual geographical scale of the scenic area and risk reduction requirements. Pick .

[0054] This method eliminates the need for iterative solutions, resulting in higher computational efficiency. It is suitable for scenic areas with straight boundaries and open terrain.

[0055] After obtaining the basic risk field, it is dynamically modulated using a pre-defined mapping model. The input to the mapping model is real-time environmental factors, including visibility data. and border control level . Data collected by the scenic area's meteorological monitoring equipment, in units of . The control level is determined by the relevant department and is represented by a discrete integer value. The mapping model is a pre-built rule mapping table, based on... The range and sum of the intervals Numerical output amplitude parameters and bandwidth parameters .

[0056] Amplitude parameters The risk values ​​of all grid points are scaled uniformly directly, and the bandwidth parameter is adjusted accordingly. By Replace with To change the extent of expansion in high-risk areas.

[0057] For example, in clear weather and under normal border control levels, Take 1.0, Take 1.0, and keep the risk field as it is; when visibility is lower than hour, The value is increased to 1.5. As the value increases to 1.2, the high-risk area expands towards both sides of the boundary, and the risk value increases at the same distance; when the border control level is raised, To increase further.

[0058] The modulated risk field is the final output risk field of S1, stored in the memory of the scenic area monitoring center server for real-time querying in subsequent steps. The risk field is stored in the form of a two-dimensional array, with each array element corresponding to the risk value of a grid point. There is a preset linear mapping relationship between the array index and the geographical coordinates of the scenic area. Through this mapping relationship, any scenic area coordinates can be converted into array indices for querying.

[0059] S2: Collect location data and calculate motion state

[0060] S2 collects tourists' location data in real time, calculates tourists' instantaneous velocity and instantaneous acceleration, and obtains the risk gradient of tourists' current location in the risk field.

[0061] As one implementation method, location data is acquired through multiple visual sensors deployed within the scenic area. The visual sensors operate at a fixed frame rate. Image acquisition, Pick That is, the time interval between two adjacent frames is .

[0062] A multi-target tracking algorithm is used to detect and correlate tourists' trajectories in images. The algorithm takes consecutive video image frames as input and outputs a unique identifier for each tourist along with their position coordinates in the image coordinate system. By using pre-calibrated camera intrinsic and extrinsic parameter matrices and incorporating a ground plane assumption, the tourist's foot points are projected onto a horizontal ground plane, converting the image coordinates into the scenic area's geographic coordinates. This method is suitable for scenic areas with deployed video surveillance systems, offering a high frequency of location updates and achieving high positioning accuracy. Magnitude.

[0063] As another implementation method, location data is reported by tourists' mobile terminals via the Global Navigation Satellite System. The scenic area server receives and parses the location message to obtain latitude and longitude coordinates, and then converts the latitude and longitude coordinates into planar coordinates under the scenic area's unified coordinate system using preset coordinate transformation parameters. This method is suitable for open scenic areas with good communication signal coverage, has low deployment costs, and the positioning accuracy is typically [insert accuracy here]. to scope.

[0064] Regardless of the method used, the measurement noise variance of the positioning system can be obtained through the equipment's nominal accuracy or through on-site static testing, denoted as... and , respectively corresponding and Noise variance in direction.

[0065] Instantaneous velocity is calculated using the three-point center difference method based on location data sequences. and instantaneous acceleration Suppose there are three consecutive sampling times. , , The corresponding positions are respectively , , Sampling interval The time difference between adjacent sampling times. This is a fixed value, determined by the data output frequency of the positioning system. Instantaneous velocity:

[0066] ;

[0067] Calculate the average speed over two time periods:

[0068] ;

[0069] ;

[0070] The acceleration can then be obtained using the following formula:

[0071] ;

[0072] At the same time, obtain the risk gradient of the tourist's current location in the risk field. By querying the risk field stored in S1, and based on the tourist's current location coordinates and the preset coordinate-index mapping relationship, the grid cell where the coordinates are located is determined. The risk value at that location is obtained by bilinear interpolation using the risk values ​​of the four corner points of that grid cell.

[0073] Again and The center difference is calculated separately for each direction, that is, the center difference is calculated along each position. and The gradient is obtained by differentiating the risk values ​​of adjacent grid points in the direction. The two components. The direction points in the direction where the risk value increases the fastest, and its magnitude indicates the degree of drastic change in the risk value at that position.

[0074] S3: Establish a force balance model and solve for the internal driving forces.

[0075] S3 establishes a force balance model:

[0076] ;

[0077] in For instantaneous acceleration, As an internal driving force, Let the motion damping force be . Substitute the known values... , , Solve And calculate based on the propagation of positioning error covariance matrix .

[0078] The risk field can be viewed as an external force field that exerts a repulsive force on tourists, moving them away from danger. The magnitude and direction of this repulsive force are determined by the risk gradient. Decision. The actual acceleration exhibited by tourists. It is its internal driving force Risk field repulsion The reaction force and the motion damping force The result of the combined force of the three. This is determined through the force balance relationship observed in the data. , , Inversely, the unobservable is deduced. .

[0079] First, calculate the motion damping force:

[0080] ;

[0081] in The damping coefficient is... direction and Same. Damping coefficient The value selection rule is as follows: a low-speed threshold is preset. , The settings are based on the location noise level and the normal walking speed of tourists. Pick .

[0082] when hour, Choose the larger first damping value , We set it to 0.8 to suppress the interference of positioning noise from tourists at low speeds or when they are stationary on the solution of driving force.

[0083] when hour, Take the smaller second damping value , Take 0.2.

[0084] Then, rearrange the terms and solve. :

[0085] ;

[0086] This calculation is a pointwise algebraic operation, performed once at each sampling time.

[0087] Next, calculations are performed based on the propagation of positioning errors. covariance matrix The noise variance determined by S2 , and sampling interval The elements in the covariance matrix of position noise propagating to velocity noise can be obtained through... , and Express.

[0088] Specifically, speed exist The noise variance in the direction is approximately ,exist The noise variance in the direction is approximately Velocity noise is then transmitted to acceleration noise through a similar differential relationship, and acceleration... exist The noise variance in direction is approximately equal to that in velocity. Directional noise variance divided by .

[0089] In the equilibrium relationship of bonding forces The linear superposition of, where The noise impact is much smaller than the positioning noise and can be ignored. The noise is caused by Noise coefficient The data is transmitted and ultimately calculated. covariance matrix . For one A symmetric matrix whose main diagonal elements are respectively exist and The variance of the direction, and the covariance of the second diagonal elements. Used for dynamic estimation of observation noise in subsequent statistical filtering.

[0090] S4: Perform orthogonal decomposition of internal driving forces

[0091] S4 with The direction is the risk direction, perpendicular to The direction is the safe direction, for Perform orthogonal decomposition to obtain the risk direction component. and the component of the safe direction .

[0092] calculate Directional unit vector:

[0093] ;

[0094] Where |g| is The modulus. When |g| is less than the preset minimum value. hour, Pick This indicates that the risk gradient at the current location is close to zero. Take the unit vector pointing from the current position to the nearest boundary point. The nearest boundary point is obtained by traversing all line segments of the border control area boundary, calculating the shortest distance from the current position to each line segment, and taking the point with the smallest distance.

[0095] Calculate perpendicular to Directional unit vector , Depend on Rotate counterclockwise Obtain. Will Projected to and Direction:

[0096] ;

[0097] ;

[0098] when This indicates that the internal driving force has a positive component in the direction of risk, that is, it propels tourists to move towards high-risk areas. The larger the value, the stronger the driving force.

[0099] when When there is no driving force pointing in the direction of risk, the tourist's current behavior is mainly driven by recreational motives focused on safety. This indicates the magnitude of the driving force that propels tourists to move in a safe direction.

[0100] Further calculate the risk tendency:

[0101] ;

[0102] Where |f| is The model. When season . This reflects the proportion of tourists whose overall driving force points to risk areas, and the value range is [value missing]. arrive .

[0103] like A lower value indicates that the driving force in the direction of risk accounts for only a small part, and tourists may have just accidentally drifted towards the boundary.

[0104] like A higher value indicates that the tourist's main intention is clearly directed towards high-risk areas.

[0105] S5: Perform statistical filtering on the risk direction component.

[0106] S5 pairs Statistical filtering is performed on the time series data to obtain the filtered risk direction component. and its estimated variance The variance of the observation noise in statistical filtering is taken as... Projection in the direction of risk.

[0107] Due to single-point solution Due to random fluctuations caused by positioning noise, directly using instantaneous values ​​for early warning decisions leads to frequent triggering and cancellation of warnings. Therefore, [the following is incomplete and requires further context:] Statistical filtering is performed on the time series data to extract stable intention trends.

[0108] As one implementation method, a Kalman filter is used for statistical filtering. The state variable of the filter is the true value of the component force in the risk direction, denoted as . .

[0109] Equations of state:

[0110] ;

[0111] in The noise is the process noise, and it follows the mean value of variance is The normal distribution, Take 0.01.

[0112] Observation equation:

[0113] ;

[0114] in The current time calculated for S4 value, To observe the noise, it follows the mean of variance is It follows a normal distribution.

[0115] Observation noise variance Pick Projected value in the direction of risk:

[0116] ;

[0117] The positioning error changes dynamically with each iteration; when the positioning error is large... Automatically increases in size, with minimal error. Automatically decreases.

[0118] The filter performs prediction and update at each sampling time.

[0119] The prediction step obtains prior estimates. Prior variance .

[0120] Update step calculates Kalman gain:

[0121] ;

[0122] Posterior estimation:

[0123] ;

[0124] Posterior variance:

[0125] ;

[0126] After filtering Pick The value, Pick The value of . Initial state Take the first time value, We set the value to 100 to reflect the high degree of uncertainty in the initial state.

[0127] As another implementation, an exponentially weighted moving average filter can be used on low-computing-power edge devices. The update formula is:

[0128] ;

[0129] in The filtered value from the previous time step. For smoothing coefficients, A value of 0.3 indicates that the weight of the new observation is 0.3 and the weight of the historical filtered value is 0.7. Take the current sliding window The sample variance is calculated, and the window length is 10 sampling points. This method requires far less computation than Kalman filtering and is suitable for scenarios with limited computing power.

[0130] S6: Cumulative Sum Detection and Fuzzy Intent Judgment

[0131] S6 performs two supplementary processes after S5.

[0132] The first step is to process... The time series data are accumulated and detected to capture mutations in tourists' risk intentions.

[0133] Set reference value Detection threshold and allowed offset Calculate the cumulative sum The initial value is .

[0134] For each new Value, calculation:

[0135] ;

[0136] like :

[0137] ;

[0138] otherwise Set as .

[0139] when Exceed At that time, the judgment A statistically significant positive jump occurs, generating an intent change marker. and Based on the preset false alarm rate requirement, and under the assumption of normal distribution, if the expected average running length is 400 sampling points, Desirable , 0.5 is acceptable.

[0140] The second step involves determining the intent in an ambiguous state. When a tourist repeatedly lingers near the boundary, their direction of movement changes frequently. Alternating positive and negative values ​​makes it difficult for the filter to converge, affecting variance estimation. With the level remaining high, tourists may be making observations and testing the waters before crossing the boundary.

[0141] Continuous monitoring Risk value of the tourist's current location , Obtained by querying the risk field stored in S1. When Continuously exceeding the preset threshold and Continuously exceeding the preset threshold Achieve continuity At each sampling time, it is determined that the tourist is in a state of ambiguous intent, and a attention prompt is generated. The filter is set according to its typical variance level under normal tracking conditions. Take the highest value of the risk field of That is, 50. Take 30, which corresponds to approximately The duration of the fluctuation is to avoid misjudgment caused by short-term fluctuations.

[0142] S7: Forward Trajectory Deduction

[0143] S7 with As a driving force, forward trajectory extrapolation is performed in conjunction with the risk field to obtain the probability of tourists entering high-risk areas in the future. and estimated arrival time The simulation is executed in real time on the scenic area's monitoring center server or edge computing devices.

[0144] As one implementation method, a stochastic deduction approach is used to determine the driving force vector:

[0145] ;

[0146] when hour Pointing in a safe direction.

[0147] Set the simulation step size Total number of steps The total simulation time is Covering future presets . There are two ways to obtain the value:

[0148] With the sampling interval in S2 Maintain consistency ,but It takes 1500 steps, which is highly accurate but computationally intensive.

[0149] Or take ,but With 600 steps, the computational efficiency is higher. The tourist's current state is their location. ,speed All are obtained from S2.

[0150] At every step of the simulation In the mean, variance is Sampled from a normal distribution This forms the driving force vector for the current step:

[0151] ;

[0152] like For negative values, the sampled mean is taken as... The variance remains the same. .

[0153] Calculate instantaneous acceleration:

[0154] ;

[0155] in Current location The risk gradient obtained in the risk field can be queried in the same way as S2. Based on the current speed The calculation method for the motion damping force is the same as that for S3.

[0156] Update speed:

[0157] ;

[0158] Update location:

[0159] ;

[0160] repeat A virtual trajectory is generated each time, resulting in a total of [number] virtual trajectories. A virtual trajectory, The value is set to 200 to ensure sufficient reliability of the statistical results. Each trajectory... Independent sampling.

[0161] The proportion of trajectories that touched the preset high-risk contour surface was obtained by statistical analysis. A high-risk equivalent surface corresponds to a risk value in the risk field that equals a preset threshold. The set of locations, A score of 90 indicates that the tourists have approached the core area of ​​the restricted zone.

[0162] If the risk value of a virtual trajectory reaches or exceeds a certain level at any step during the simulation period... If the initial contact time is not reached, it is considered a contact. The average of the first contact time is calculated from the contact trajectory. If no trajectory is touched, then Take the preset maximum value , Pick This indicates that under the current driving force assumptions, tourists will not enter high-risk areas.

[0163] As another implementation method, deterministic deduction can be used when computing resources are scarce, with a constant driving force at each step. This generates a deterministic trajectory. Upon contact... , First touch time, otherwise , Deterministic extrapolation involves less computation, but it does not fully account for the impact of driving force uncertainties on the trajectory.

[0164] S8: Tiered Early Warning

[0165] S8 based on and The system outputs tiered early warnings, which are transmitted through the scenic area's monitoring system display terminal, security personnel's handheld terminals, or the scenic area's broadcasting system.

[0166] when and If the time is right, it is considered safe and no warning is triggered.

[0167] when and At that time, a yellow alert will be issued. This represents the shortest time required for security personnel to arrive at the scene and implement intervention after receiving a warning, based on the scenic area's emergency plan and geographical layout. Pick Once a yellow alert is issued, security personnel can arrive in advance to deploy near the alert location, and the system will issue boundary safety reminders to tourists in the area through scenic area broadcasts or information push notifications.

[0168] when and At that time, a red alert is output, and the following information is given: Orientation guidance is provided in the opposite direction to the current visitor's location. This guidance is generated by taking the risk gradient at the current visitor's location. Calculate its opposite unit vector This direction represents the fastest decrease in risk value. After comparison with the scenic area's geographic orientation data, it is converted into a readable directional description and provided to security personnel, for example, "approximately northeast." "Intercept at the location."

[0169] When S6 generates an intent change flag and At that time, a red alert and directional guidance will be issued. This rule is used to detect high-risk situations with the intention to change course: even if Not achieved However, the tourists suddenly showed a clear intention to move closer to high-risk areas, which, combined with a certain probability of collision, constituted a condition that required immediate intervention.

[0170] The above threshold Take 0.1, Take 0.7, Take 0.5.

[0171] As one implementation method, tiered early warning can also be combined with risk directionality. Integrate and adjust the intent ambiguity status markers. When in an intent ambiguity state and Between and During this period, the attention level will be increased, and a prompt message will be sent to the monitoring personnel. The message will include the visitor's current location and the duration of the ambiguous state, facilitating manual analysis by the monitoring personnel. High value and Even when it continues to increase, Greater than This can also trigger a yellow alert in advance, increasing the margin for safe handling.

[0172] Through the aforementioned steps S1 to S8, this application analyzes the risk tendency in the internal driving force of tourists based solely on early fragments of their movement trajectories before they enter high-risk areas, and achieves proactive prevention of boundary crossing risks through forward extrapolation and graded early warning.

[0173] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent process changes made using the content of this application specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for early warning of safety risks in border tourism, characterized in that, include: S1: For border control zones without physical isolation in border tourist areas, generate a risk field that changes with real-time environmental data. The risk field assigns a continuous risk value to each location within the area. S2: Collect tourists' location data in real time, calculate tourists' instantaneous velocity and instantaneous acceleration, and obtain the risk gradient of tourists' current location in the risk field; S3: Establish a force balance model, and express the instantaneous acceleration as the resultant force of the internal driving force, the reaction force of the risk gradient, and the motion damping force. Substitute the known instantaneous acceleration, the risk gradient, and the motion damping force to solve for the internal driving force. S4: Using the direction of the risk gradient as the risk direction and the direction perpendicular to the risk gradient as the safety direction, orthogonally decompose the internal driving force to obtain the risk direction component force and the safety direction component force; S5: Perform statistical filtering on the time series of the risk directional force to obtain the filtered risk directional force and its estimated variance; S6: Using the filtered risk direction component as the driving force, and combining it with the risk field, forward trajectory deduction is performed to obtain the probability of tourists entering high-risk areas in the future and the expected arrival time. S7: Output a graded warning based on the probability and the expected arrival time.

2. The border tourism safety risk early warning method according to claim 1, characterized in that, The risk field described in step S1 is generated as follows: Using the boundary of the border control zone as the source, an anisotropic diffusion method is used to generate a basic risk field, wherein the diffusion coefficient along the perpendicular direction of the boundary is greater than the diffusion coefficient along the parallel direction of the boundary. The basic risk field is modulated by using a preset mapping model to output amplitude and bandwidth parameters based on the input real-time environmental factors, thereby obtaining the risk field.

3. The border tourism safety risk early warning method according to claim 2, characterized in that, The real-time environmental factors include visibility data and border control level data.

4. The border tourism safety risk early warning method according to claim 1, characterized in that, The motion damping force mentioned in step S3 is obtained by multiplying the damping coefficient by the instantaneous velocity. The value of the damping coefficient is related to the magnitude of the instantaneous velocity. When the instantaneous velocity is lower than a preset low-speed threshold, the value of the damping coefficient increases.

5. The border tourism safety risk early warning method according to claim 1, characterized in that, Step S3 also includes calculating the covariance matrix of the internal driving force based on the propagation of the positioning error; In step S5, the observation noise variance of the statistical filtering is taken as the projection of the covariance matrix onto the risk direction.

6. The border tourism safety risk early warning method according to claim 5, characterized in that, The statistical filtering described in step S5 is implemented using a Kalman filter, with the risk direction component as the observation value and the random walk process as the state equation.

7. The border tourism safety risk early warning method according to claim 1, characterized in that, Step S5 is followed by: The time series of the filtered risk direction component is accumulated and detected. When a statistically significant positive jump is detected in the filtered risk direction component, an intention change marker is generated.

8. The border tourism safety risk early warning method according to claim 1, characterized in that, Step S5 is followed by: When the estimated variance is consistently higher than a preset first threshold, and the risk value of the tourist's current location is higher than a preset second threshold, the tourist is determined to be in a state of ambiguous intent, and a attention prompt is generated.

9. The border tourism safety risk early warning method according to claim 1, characterized in that, The method for forward trajectory deduction in step S6 is as follows: Using the tourist's current position and instantaneous speed as the initial state, and the filtered risk direction component and the estimated variance as driving force parameters, multiple virtual trajectories are generated in the risk field. The probability is calculated as the proportion of the virtual trajectory that touches a preset high-risk contour surface, and the average touching time is calculated as the expected arrival time.

10. The border tourism safety risk early warning method according to claim 7 or 9, characterized in that, The graded early warning mentioned in step S7 includes: When the probability is lower than a preset lower limit and the filtered risk directional force is less than or equal to zero, it is determined to be safe and no warning is triggered. When the probability exceeds a preset upper limit and the expected arrival time is greater than a preset safety intervention time, a yellow warning is issued; When the probability exceeds the preset upper limit and the estimated arrival time is less than or equal to the safety intervention time, a red alert is output, and directional guidance is given to intervene in the opposite direction of the risk gradient. When the intent change marker is generated and the probability exceeds a preset threshold, a red alert is output, and directional guidance is given to intervene in the opposite direction of the risk gradient.

Images