A vehicle crash detection and intervention method and system

Abstract

The application provides a vehicle collision detection and intervention method and system, and relates to the technical field of data processing.The method comprises the following steps: based on a minimum safety distance, performing spatial modeling analysis on a dynamically selected vehicle motion trajectory curvature center point, a vehicle safety envelope boundary point and a historical trajectory weighted center point to obtain a geometric feature model; converting the geometric feature model into a corresponding virtual feature point set, and constructing a dynamic evaluation structure based on the virtual feature point set; based on the dynamic evaluation structure, performing safety threat level calculation to obtain a vehicle safety threat index; determining a warning level according to the vehicle safety threat index to obtain a warning level; and controlling a roadside device to perform corresponding warning and intervention operations according to the warning level.The application can identify the trend of vehicle deviation from the normal driving path in advance through trajectory prediction, and shorten the time window of risk response.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for vehicle collision detection and intervention. Background Technology

[0002] In recent years, the requirements for real-time and accurate traffic safety management in urban roads and specific controlled areas have been increasing; traditional vehicle monitoring technologies have the following limitations:

[0003] For example, many existing systems rely on a single sensor data source (such as pure vision or pure radar technology), which lacks sufficient perception stability under complex lighting and weather conditions, making it difficult to achieve continuous and accurate tracking of vehicle motion. Common trajectory prediction methods are mostly based on historical averages or simple linear extrapolation, failing to fully consider vehicle kinematic characteristics and real-time heading changes, resulting in significant prediction errors. Safety risk assessments are often one-dimensional, typically relying only on vehicle speed or location information, lacking multi-parameter fusion analysis of trajectory deviation, dynamic safety distance, and other parameters, making it difficult to support tiered early warning and precise intervention.

[0004] In addition, some existing roadside equipment control strategies lag behind the actual occurrence of risks, failing to achieve a closed-loop linkage of perception, assessment, and intervention. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a vehicle collision detection and intervention method and system, which can identify the trend of a vehicle deviating from the normal driving path in advance through trajectory prediction, thereby shortening the risk response time window.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] A vehicle collision detection and intervention method, the method comprising:

[0008] The system can identify static feature data and dynamic operating parameters of vehicles within the controlled area in real time. The static feature data includes license plate, license plate type, vehicle type, and vehicle brand. The dynamic operating parameters include vehicle speed, acceleration, heading angle, and position coordinates.

[0009] Based on dynamic operating parameters, motion trajectory fitting and prediction are performed, and vehicle trajectory deviation is calculated.

[0010] Based on the vehicle trajectory deviation and the aforementioned dynamic operating parameters, a collision risk assessment is performed, and the minimum safe distance between the vehicle and the boundary is calculated.

[0011] Based on the minimum safety distance, spatial modeling analysis is performed on the dynamically selected vehicle trajectory curvature center point, vehicle safety envelope boundary point, and historical trajectory weighted center point to obtain a geometric feature model; the geometric feature model is converted into a corresponding virtual feature point set, and a dynamic evaluation structure is constructed based on the virtual feature point set;

[0012] Based on the aforementioned dynamic evaluation structure, the security threat level is calculated to obtain the vehicle security threat index;

[0013] The warning level is determined based on the vehicle safety threat index.

[0014] Based on the warning level, control the roadside equipment to perform corresponding warning and intervention operations.

[0015] A vehicle collision detection and intervention system, comprising:

[0016] The data acquisition module is used to identify the static feature data and dynamic operating parameters of vehicles within the controlled area in real time. The static feature data includes license plate, license plate type, vehicle type, and vehicle brand; the dynamic operating parameters include vehicle speed, acceleration, heading angle, and position coordinates.

[0017] The calculation module is used to fit and predict the motion trajectory based on dynamic operating parameters, and calculate the vehicle trajectory deviation; based on the vehicle trajectory deviation and the dynamic operating parameters, it performs a collision risk assessment and calculates the minimum safe distance between the vehicle and the boundary.

[0018] The module is used to perform spatial modeling analysis on the dynamically selected vehicle trajectory curvature center point, vehicle safety envelope boundary point, and historical trajectory weighted center point based on the minimum safety distance to obtain a geometric feature model; the geometric feature model is converted into a corresponding virtual feature point set, and a dynamic evaluation structure is constructed based on the virtual feature point set;

[0019] The early warning module is used to calculate the safety threat level based on the dynamic evaluation structure to obtain the vehicle safety threat index; determine the early warning level based on the vehicle safety threat index to obtain the early warning level; and control the roadside equipment to perform corresponding early warning and intervention operations based on the early warning level.

[0020] The above-described solution of the present invention has at least the following beneficial effects:

[0021] By collecting real-time static vehicle feature data (license plate, type, model, brand) and dynamic operating parameters (vehicle speed, acceleration, etc.), the system compensates for the incomplete data collection of traditional systems and reduces the risk of misjudgment or omission due to missing key information.

[0022] By fitting and predicting trajectories based on dynamic parameters and calculating deviation, the tendency of vehicles to deviate from their normal paths can be identified as early as possible, allowing sufficient time for early warning and intervention, thus solving the problem of delayed early warning in traditional security systems. By combining trajectory deviation with dynamic parameters to calculate the minimum safe distance, abstract collision risks are transformed into quantifiable indicators, replacing traditional subjective experience judgments and avoiding judgment biases caused by environmental interference, making risk assessment more accurate. By spatially modeling the trajectory curvature center and safety envelope boundary and constructing a dynamic evaluation structure, the system fully considers vehicle model differences, motion trends, and historical behavior, adapting to different vehicle models and traffic conditions, thus breaking through the limitations of traditional single-dimensional evaluation. Attached Figure Description

[0023] Figure 1 This is a schematic flowchart of a vehicle collision detection and intervention method provided by an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of a vehicle collision detection and intervention system provided in an embodiment of the present invention. Detailed Implementation

[0025] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0026] like Figure 1 As shown, an embodiment of the present invention proposes a vehicle collision detection and intervention method, the method comprising the following steps:

[0027] Step S1: Identify the static feature data and dynamic operating parameters of vehicles within the controlled area in real time. The static feature data includes license plate, license plate type, vehicle type, and vehicle brand. The dynamic operating parameters include vehicle speed, acceleration, heading angle, and position coordinates.

[0028] Step S2: Based on dynamic operating parameters, perform motion trajectory fitting and prediction, and calculate the vehicle trajectory deviation.

[0029] Step S3: Based on the vehicle trajectory deviation and the dynamic operating parameters, a collision risk assessment is performed, and the minimum safe distance between the vehicle and the boundary is calculated.

[0030] Step S4: Based on the minimum safety distance, perform spatial modeling analysis on the dynamically selected vehicle trajectory curvature center point, vehicle safety envelope boundary point, and historical trajectory weighted center point to obtain a geometric feature model; convert the geometric feature model into a corresponding virtual feature point set, and construct a dynamic evaluation structure based on the virtual feature point set;

[0031] Step S5: Based on the dynamic evaluation structure, calculate the security threat level to obtain the vehicle security threat index;

[0032] Step S6: Determine the warning level based on the vehicle safety threat index to obtain the warning level;

[0033] Step S7: Based on the warning level, control the roadside equipment to perform corresponding warning and intervention operations.

[0034] In this embodiment of the invention, trajectory fitting and deviation calculation can identify unexpected vehicle movement trends in advance (e.g., a vehicle that should be traveling along the lane but deviates towards the boundary of the controlled area (guardrail, sidewalk)). Compared with the traditional mode that only issues a warning when the vehicle approaches the boundary, this can detect risks in advance and allow sufficient time for intervention. The quantitative calculation of the minimum safe distance transforms the collision risk from an ambiguous yes or no to a specific distance value, avoiding errors caused by subjective human judgment (e.g., misjudging the safe distance by visual observation), making risk assessment more objective and quantifiable. The geometric feature model, combined with the spatial geometric relationship of the vehicle's real-time movement (e.g., trajectory curvature reflects the vehicle's turning radius, and the safety envelope boundary point reflects the vehicle's maximum occupied space), avoids the lag of static models using fixed parameters to evaluate dynamic vehicles (e.g., when the vehicle accelerates, the static model still evaluates the safe distance based on the original speed). The dynamic evaluation structure constructed by the virtual feature point set can be updated in real time with the vehicle's movement (e.g., when the vehicle's position coordinates change, the curvature center point and weighted center point are adjusted synchronously), ensuring that the risk assessment always matches the actual movement state of the vehicle, reducing the probability of misjudging low risk as high risk (over-warning) or misjudging high risk as low risk (insufficient warning). Low-level warnings may trigger only minor interventions such as roadside warning lights and voice announcements, while high-level warnings may trigger major interventions such as emergency braking guidance (e.g., roadside crash barrier warnings and ESP-linked vehicle alerts), avoiding resource waste and improving intervention efficiency. Roadside equipment (such as smart barriers, LED warning screens, voice broadcasts, and millimeter-wave radar guidance devices) directly respond to the warning level, achieving automatic intervention (without manual operation) and reducing the time lag of manual response.

[0035] In a preferred embodiment of the present invention, step S1 involves real-time identification of static feature data and dynamic operating parameters of vehicles within the controlled area. The static feature data includes license plate number, license plate type, vehicle model, and vehicle brand. The dynamic operating parameters include vehicle speed, acceleration, heading angle, and position coordinates.

[0036] Step S101 involves collecting multi-modal vehicle data using a multi-source sensor array deployed at key locations within the controlled area. This array includes visual sensors, millimeter-wave radar, and lidar. Specifically, in the controlled area, such as at school gates, entrances and exits of controlled road sections, and densely populated intersections, the multi-source sensor array is deployed according to the principles of comprehensive coverage and complementary data. The visual sensors, specifically high-definition industrial cameras, are installed at high points within the controlled area, such as light poles and monitoring poles, to ensure coverage of the entire vehicle's travel path. Millimeter-wave radar is deployed on both sides of the road or in medians to adapt to adverse weather conditions such as rain, snow, and fog. LiDAR is installed at the entrance and core road sections of the controlled area to acquire high-precision 3D point cloud data. A time synchronization module, such as GPS or a network clock, aligns the acquisition frequencies of the visual sensors, millimeter-wave radar, and lidar with the timestamps, ensuring that the three types of sensors collect target vehicle data in the same time dimension. The visual sensors acquire vehicle exterior images, the millimeter-wave radar acquires information such as vehicle speed and distance, and the lidar acquires the vehicle's 3D contour and precise location point cloud, forming a multi-modal raw data set.

[0037] Step S102 involves preprocessing the vehicle images acquired by the vision sensor to obtain optimized vehicle image data. Specifically, this includes using noise reduction algorithms, such as Gaussian filtering and median filtering, to process the images, removing irrelevant noise and retaining key information such as vehicle outlines and license plates, for possible noise in the original vehicle images acquired by the vision sensor, such as raindrops on rainy days, light interference at night, and camera noise.

[0038] Step S103 involves extracting and recognizing vehicle features based on the optimized vehicle image data. This includes parsing license plate characters, license plate type, vehicle model, and vehicle brand information to form a static feature dataset. Specifically, this includes: using object detection algorithms to locate the license plate region based on the optimized vehicle images (e.g., using YOLO, SSD, etc. to identify the license plate position); then using character segmentation techniques, such as projection segmentation, to separate the license plate characters; and finally, using character recognition models, such as CNN convolutional neural networks, to parse the license plate characters. Simultaneously, by comparing with a license plate database, such as public security traffic management vehicle license plate registration information, the legality of the license plate is determined, distinguishing between fake, cloned, and legitimate plates, and recording the license plate number and legality indicator. Based on the license plate color (e.g., blue for small vehicles, yellow for large vehicles, green for new energy vehicles) and character format (e.g., D / F suffix for new energy vehicle license plates, and hanging characters for large vehicle license plates), combined with a pre-defined license plate type rule base, the vehicle's license plate type is determined (e.g., small car license plate, large truck license plate, new energy vehicle license plate). Finally, through object detection and image feature matching techniques, the vehicle body contour, such as wheelbase, is extracted. The system identifies vehicle height, presence of a cargo box, and key component features such as headlight shape, grille style, and wheel design. These features are compared against a vehicle model database containing exterior feature libraries for different brands and models to determine the vehicle type (e.g., sedan, SUV, truck, bus) and brand (e.g., a specific brand of sedan or truck). Information such as license plate number, license plate legality, license plate type, vehicle model, and brand is then integrated to assign a unique identifier (ID) to each identified vehicle, forming a static feature dataset containing the vehicle ID, static feature items, and feature values.

[0039] Step S104 involves performing point cloud registration and target clustering processing based on multi-frame point cloud data synchronously acquired by millimeter-wave radar and lidar to establish spatiotemporal trajectory information of the vehicle target. Specifically, this includes: Since the acquisition frequencies of millimeter-wave radar and lidar may differ (e.g., millimeter-wave radar acquires 10 frames per second, lidar acquires 20 frames per second), based on the time synchronization module in step S101, aligning the multi-frame point cloud data of the two types of radars by timestamp to ensure that radar data at the same time point corresponds to the same vehicle motion state; Since the installation positions and detection angles of millimeter-wave radar and lidar are different, converting the point cloud data of the two types of radars to the same world coordinate system, such as a coordinate system established with a fixed point in the controlled area as the origin, and using a point cloud registration algorithm, such as the ICP iterative nearest point algorithm, to eliminate the differences between the two systems. The coordinate system deviation of radar-like systems is used to achieve spatial fusion of point cloud data. For the fused multi-frame point cloud data, clustering algorithms, such as DBSCAN density clustering and K-Means clustering, are used to classify the point clouds. Point clouds belonging to the same vehicle are clustered into a target cluster, excluding background point clouds such as roadside guardrails, trees, and streetlights. At the same time, different vehicles are initially distinguished by the size and shape of the clusters to avoid misclassifying the point clouds of multiple adjacent vehicles as the same target. Based on the clustering results of the multi-frame point clouds, the center coordinates of each vehicle cluster at each timestamp are recorded, i.e., the approximate position of the vehicle. The position coordinates of the same vehicle in multiple frames are concatenated in chronological order. By the positional continuity and shape consistency of the clusters, it is determined whether they belong to the same vehicle, forming the vehicle spatiotemporal trajectory information containing vehicle ID, timestamp, and position coordinates.

[0040] Step S105: Based on the spatiotemporal trajectory information of the vehicle targets, determine the real-time position, motion contour, and attitude information of each vehicle target. Specifically, this includes: extracting the center coordinates of each vehicle cluster at the current timestamp from the spatiotemporal trajectory information generated in step S104, or using the vehicle centroid coordinates fitted by the LiDAR point cloud, as the real-time position coordinates of the vehicles to ensure that the position information is synchronized with the current time; for the LiDAR point cloud clusters of vehicles at the current timestamp, using geometric fitting algorithms, such as the minimum bounding rectangle algorithm and the convex hull algorithm, to fit the point cloud boundary, determine the vehicle's body length, width, height, and other dimensional parameters, and form the vehicle's motion contour, i.e., the spatial occupancy range of the vehicle at the current position; by analyzing the vehicle's spatiotemporal trajectory coordinates in two adjacent frames, i.e., the current frame and the previous frame, calculating the direction of change of the vehicle's position, i.e., the initial direction of the vehicle's travel; and simultaneously, combining the orientation of the vehicle's motion contour fitted by the LiDAR point cloud, such as the orientation of the long side of the vehicle body, correcting the vehicle's travel direction, and determining the vehicle's real-time attitude information, such as the attitude being 0° when the vehicle is traveling straight and 30° to the left when turning left.

[0041] Step S106 involves fusing the spatiotemporal trajectory information of the vehicle target with its real-time position, motion profile, and attitude information, and performing motion state estimation to obtain the vehicle's velocity, acceleration, and heading angle parameters, thus forming a dynamic operating parameter dataset. Specifically, this includes associating the vehicle's spatiotemporal trajectory information from step S104 with the real-time position, motion profile, and attitude information from step S105. Using the vehicle's unique identifier ID as a link, this ensures that the trajectory, position, profile, and attitude information of the same vehicle correspond and match, eliminating data redundancy or misalignment, such as avoiding confusion between the trajectory of vehicle A and the attitude of vehicle B.

[0042] Based on the real-time vehicle position coordinates of two adjacent frames, the distance between the two frames, i.e., the spatial distance, is calculated and divided by the time interval between the two frames, such as 0.1 seconds, to obtain the instantaneous speed of the vehicle. If there is data from multiple frames, the vehicle speed calculation result can be optimized by using the sliding window averaging method to reduce fluctuations. Based on the instantaneous speed of two adjacent timestamps, the speed difference is calculated and divided by the time interval between the two timestamps to obtain the vehicle's acceleration. Positive values ​​indicate acceleration, and negative values ​​indicate deceleration. At the same time, by judging the magnitude of the absolute value of the acceleration, normal acceleration is distinguished from rapid acceleration, and normal deceleration is distinguished from rapid deceleration. Combining the vehicle attitude information from step S105, i.e., the driving direction, and the curvature change of the spatiotemporal trajectory, such as the tangent direction of the trajectory when turning, the vehicle heading angle is corrected. Using the coordinate system of the controlled area as a reference, the vehicle's driving direction is converted into a standard heading angle, such as 0° for true north and 90° for true east, to ensure that the heading angle accurately reflects the vehicle's driving direction. Parameters such as vehicle speed, acceleration, heading angle, and real-time position coordinates of the same vehicle are organized into a dynamic operation parameter dataset according to the format of vehicle ID, dynamic parameter items, parameter values, and timestamps. By establishing the vehicle's spatiotemporal trajectory through point cloud registration and target clustering, the vehicle's spatiotemporal trajectory can be accurately distinguished from background targets such as roadside guardrails, avoiding target confusion and ensuring trajectory continuity and accuracy.

[0043] In a preferred embodiment of the present invention, step S2, which involves fitting and predicting the motion trajectory based on dynamic operating parameters and calculating the vehicle trajectory deviation, includes:

[0044] Step S201: Based on the dynamic operating parameter dataset, extract the position coordinate data of the vehicle's continuous time series and construct a vehicle motion trajectory point set. Specifically, this includes: extracting the real-time position coordinate data of each vehicle from the dynamic operating parameter dataset generated in step S106, and simultaneously associating the timestamp information of the corresponding data; sorting the position coordinate data according to the order of the timestamps to ensure a continuous time series and avoid trajectory breakage due to disordered time order; during this process, removing abnormal position coordinates caused by temporary sensor failures or external interference, such as coordinates that exceed the reasonable range of the control area or coordinates that have significant jumps with adjacent time point coordinates, and finally constructing a vehicle motion trajectory point set containing timestamps, vehicle IDs, and position coordinates.

[0045] Step S202: Based on the vehicle motion trajectory point set, determine the order of the polynomial function used for fitting; construct the coefficient matrix and constant term matrix of the polynomial fitting based on the time series coordinates of the trajectory point set; calculate the least squares solution of the coefficients of each term of the polynomial by solving the canonical equation formed by the coefficient matrix and the constant term matrix; determine the final fitting polynomial coefficients based on the least squares solution to obtain the actual driving trajectory model characterizing the continuous position change law of the vehicle, specifically including:

[0046] Based on the dynamic operation parameter dataset output in step S106, filter the relevant data of the target vehicle by the vehicle's unique identifier ID, extract the vehicle's position coordinates (including x-axis horizontal coordinates and y-axis vertical coordinates) and corresponding timestamp information under the continuous time series, and form an initial trajectory point set, which may contain a combination of multiple times, horizontal coordinates and vertical coordinates.

[0047] Because multi-source sensors (millimeter-wave radar, lidar) may be subject to external interference during the data acquisition process (such as point cloud occlusion caused by trees around the campus or pedestrians crossing), the initial trajectory point set needs to be cleaned: if the spatial distance between a trajectory point and two adjacent trajectory points exceeds the instantaneous displacement range of normal vehicle travel within the controlled area (usually set to 0.3-0.8 meters in combination with the speed limit of roads around the campus), or if the time interval between the timestamp of the point and the adjacent points exceeds twice the fixed acquisition cycle of the sensor (e.g., lidar acquires 10 frames per second, and an interval exceeding 0.2 seconds is considered abnormal), then the point is marked as abnormal data and deleted, finally obtaining a clean trajectory point set without interference and continuous.

[0048] Based on the actual needs of the road scenarios around the campus (including straight roads, gentle curves, and sharp bends at intersections), the polynomial order is determined by quantifying the directional changes of trajectory points to avoid subjective judgment bias.

[0049] First, calculate the driving direction of every two consecutive points in the cleaning trajectory point set. For any two adjacent trajectory points (such as the i-th point and the (i+1)-th point), determine the driving direction of that segment by the direction of the line connecting the two points (e.g., using the positive x-axis as a reference, described as deviating 3° due north, 15° due east, etc.). Calculate the angle difference between the driving directions of two adjacent segments (e.g., the angle between the direction from point (i-1) to point i and the direction from point i to point (i+1)), and take the average of all differences. If the average direction change is ≤ 5° (corresponding to straight road segments around the campus, the trajectory...), then... If the trajectory is close to a straight line, use a 1st or 2nd order polynomial (1st order is suitable for uniform speed straight driving, 2nd order is suitable for uniform acceleration / uniform deceleration straight driving) to avoid overfitting of the trajectory caused by higher order polynomials; if the average direction change is between 5° and 30° (corresponding to gentle curves around the campus, where the trajectory is slightly curved), use a 3rd order polynomial to balance fitting accuracy and model complexity; if the average direction change is >30° (corresponding to sharp curves at campus intersections and lane-changing scenarios, where the trajectory is significantly curved), use a 4th order polynomial to ensure accurate capture of the trajectory curvature changes.

[0050] Meanwhile, to verify the rationality of the order, the clean trajectory point set is divided into two parts: 70% is used as the training point set to fit the polynomial, and 30% is used as the verification point set. The fitted polynomial is applied to the verification point set. If the deviation between the actual coordinates of the verification point set and the fitted coordinates exceeds 0.1 meters (which meets the requirements of lidar positioning accuracy), the order is readjusted and the polynomial is fitted again until the deviation meets the requirements.

[0051] Constructing a polynomial fitting matrix along the axes (independent operations on the x and y axes):

[0052] Since the vehicle's position within the controlled area needs to be described by both the x-axis (lateral, such as the left-right direction of the road) and the y-axis (vertical, such as the front-back direction of the road) coordinates, and the motion patterns of the two axes are independent, fitting matrices need to be constructed separately:

[0053] Construction of the x-axis direction matrix, for the x-axis horizontal coordinates and corresponding timestamps of the clean trajectory point set:

[0054] Construct a coefficient matrix. The number of rows in the coefficient matrix is ​​the same as the number of points in the clean trajectory point set, and the number of columns is related to the order of the polynomial (e.g., a 3rd order polynomial corresponds to 4 columns, corresponding to the 0th, 1st, 2nd, and 3rd powers of the timestamp, respectively). For example, if a timestamp is t1, then the row data corresponding to this timestamp in the coefficient matrix is ​​1 (t1 to the power of 0), t1 (t1 to the power of 1), t1×t1 (t1 to the power of 2), and t1×t1×t1 (t1 to the power of 3). All timestamps are filled in sequentially according to the above rules to form a complete x-axis coefficient matrix. Construct a constant term matrix. The number of rows in the constant term matrix is ​​the same as the number of points in the clean trajectory point set, and it contains only 1 column. The value of each row corresponds to the x-axis horizontal coordinate of that timestamp in the clean trajectory point set, and they are arranged sequentially according to the timestamp order.

[0055] Construction of the y-axis direction matrix, for the vertical y-axis coordinates and corresponding timestamps of the clean trajectory point set:

[0056] The construction rules for the coefficient matrix are consistent with those for the x-axis. The number of rows, columns, and timestamp powers are filled in exactly the same way (to ensure synchronization of the time dimension). The constant term matrix contains only one column. The value of each row corresponds to the vertical y-axis coordinate of that timestamp in the clean trajectory point set, and they are arranged in the order of the timestamps.

[0057] To solve the canonical equation and verify the polynomial coefficients, we first construct the canonical equation:

[0058] The two key components of the regular equation are determined as follows: a 4x4 square matrix (obtained by multiplying the transpose of the x-axis coefficient matrix by the original coefficient matrix) and a 4x1 column matrix (obtained by multiplying the transpose of the x-axis coefficient matrix by the x-axis constant term matrix). The x-axis coefficient matrix is ​​a matrix with n rows and 4 columns representing the number of clean trajectory points (denoted as n). Since it fits a 3rd-order polynomial, the columns correspond to the 0th, 1st, 2nd, and 3rd powers of the timestamps. The x-axis constant term matrix is ​​an nx1 matrix (the columns correspond to the actual x-axis coordinates of each clean trajectory point).

[0059] Calculate the square matrix of the regular equation (transpose of the coefficient matrix × original coefficient matrix):

[0060] Step 1: Transpose the x-axis coefficient matrix:

[0061] Transform the original n x-axis coefficient matrix into a 4 x n transpose matrix. The specific transpose rules are as follows: the first row of the transpose matrix corresponds exactly to the first column of the original coefficient matrix (i.e., all timestamps raised to the power of 0; since any number raised to the power of 0 is 1, all n elements in the first row of the transpose matrix are 1); the second row of the transpose matrix corresponds exactly to the second column of the original coefficient matrix (i.e., all timestamps raised to the power of 1, which is the value of each timestamp itself); the third row of the transpose matrix corresponds exactly to the third column of the original coefficient matrix (i.e., all timestamps raised to the power of 2, which is the result of multiplying each timestamp value by itself); and the fourth row of the transpose matrix corresponds exactly to the fourth column of the original coefficient matrix (i.e., all timestamps raised to the power of 3, which is the result of multiplying the square of each timestamp value by the timestamp value).

[0062] Step 2: Multiply the transpose matrix by the original coefficient matrix to obtain a square matrix:

[0063] Multiplying the 4x4 transpose matrix with the original nx4 coefficient matrix yields a 4x4 square matrix. The calculation method for each element must be determined row by row and column by column.

[0064] For the elements in the first row and first column of the square matrix, take the n elements (all 1s) of the first row of the transpose matrix and multiply them one-to-one with the n elements (all 1s) of the first column of the original coefficient matrix (each product is 1×1=1). Then add these n products together. The result is equal to the number of clean trajectory points n. For the elements in the first row and second column of the square matrix, take the n elements (all 1s) of the first row of the transpose matrix and multiply them one-to-one with the n elements (timestamp raised to the power of 1) of the second column of the original coefficient matrix (each product is 1×timestamp value). Then add these n products together. The result is equal to the sum of all timestamp values. For the elements in the first row and third column of the square matrix, take the n elements (all 1s) of the first row of the transpose matrix and multiply them one-to-one with the n elements (timestamp raised to the power of 2) of the third column of the original coefficient matrix (each product is 1×timestamp value squared). Then add these n products together. The result is equal to the sum of the squares of all timestamp values.

[0065] For the elements in the first row and fourth column of the square matrix, take the n elements (all 1s) of the first row of the transpose matrix, and multiply them one-to-one with the n elements (timestamps raised to the power of 3) of the fourth column of the original coefficient matrix (each product is 1 × the cube of the timestamp value). Then sum these n products; the result is equal to the sum of the cubes of all timestamp values. For the elements in the second row and first column of the square matrix, take the n elements (timestamps raised to the power of 1) of the second row of the transpose matrix, and multiply them one-to-one with the n elements (all 1s) of the first column of the original coefficient matrix (each product is 1 × the cube of the timestamp value). The product is timestamp value × 1. Then, add these n products together. The result is the same as the element in the first row and second column of the square matrix (which is the sum of all timestamp values). The element in the second row and second column of the square matrix is ​​taken as the n elements in the second row of the transpose matrix (timestamp raised to the power of 1). Each of these n elements is multiplied one by one with the n elements in the second column of the original coefficient matrix (timestamp raised to the power of 1). Each product is timestamp value × timestamp value = timestamp value squared. Then, add these n products together. The result is equal to the sum of the squares of all timestamp values.

[0066] The elements in the 2nd row and 3rd column of the square matrix are taken from the n elements (timestamps raised to the power of 1) of the 2nd row of the transpose matrix, and multiplied one-to-one with the n elements (timestamps raised to the power of 2) of the 3rd column of the original coefficient matrix (each product is timestamp value × timestamp value squared = timestamp value cubed). These n products are then summed, and the result is equal to the sum of the cubes of all timestamp values. The elements in the 2nd row and 4th column of the square matrix are taken from the n elements (timestamps raised to the power of 1) of the 2nd row of the transpose matrix, and multiplied one-to-one with the n elements (timestamps raised to the power of 2) of the 3rd column of the original coefficient matrix. The n elements (timestamps raised to the power of 3) in the 4th column of the coefficient matrix are multiplied one by one (each product is timestamp value × timestamp value cubed = timestamp value raised to the power of 4). These n products are then added together, and the result is equal to the sum of all timestamp values ​​raised to the power of 4. By doing this, following the rule of multiplying the elements of a row of the transpose matrix with the elements of a column of the original coefficient matrix and then summing them, all the elements in the remaining 3rd and 4th rows of the square matrix are calculated row by row and column by column, and finally the 4x4 square matrix is ​​constructed.

[0067] Calculate the column matrix of the regular equation (coefficient matrix transpose × constant term matrix):

[0068] Step 1, determine the structure of the x-axis constant term matrix:

[0069] The x-axis constant term matrix is ​​an n-row, 1-column matrix. Each row of elements corresponds to the actual x-axis coordinate of a clean trajectory point, and the order of the elements is completely consistent with the order of the timestamps of the clean trajectory points (i.e., the first row corresponds to the actual x-axis coordinate of the first timestamp, the second row corresponds to the actual x-axis coordinate of the second timestamp, and so on, up to the nth row corresponding to the actual x-axis coordinate of the nth timestamp).

[0070] Step 2: Multiply the transpose matrix by the constant term matrix to obtain the column matrix:

[0071] Multiply the 4xn transpose matrix (the same transpose matrix used when calculating square matrices) with the nx1 xx axis constant term matrix to obtain a 4x1 column matrix; the calculation method for each element is as follows:

[0072] The elements of the first row of the column matrix are taken as follows: Take the n elements (all 1s) of the first row of the transpose matrix, and multiply them one-to-one with the n elements (actual x-coordinates) of the first column of the constant term matrix (each product is 1 × actual x-coordinate = actual x-coordinate). Then sum these n products; the result is equal to the sum of the actual x-coordinates of all clean trajectory points. The elements of the second row of the column matrix are taken as follows: Take the n elements (timestamps raised to the power of 1) of the second row of the transpose matrix, and multiply them one-to-one with the n elements (actual x-coordinates) of the first column of the constant term matrix (each product is timestamp value × actual x-coordinate). Then sum these n products; the result is equal to the sum of the products of all timestamp values ​​and their corresponding actual x-coordinates. The elements of the third row of the column matrix are taken as follows: Take the n elements (timestamps raised to the power of 1) of the third row of the transpose matrix... The timestamp (power of 2) is multiplied one-to-one with each of the n elements (actual x-coordinates) in the first column of the constant term matrix (each product is the square of the timestamp value × the actual x-coordinate). These n products are then summed, and the result is equal to the sum of the products of the squares of all timestamp values ​​and their corresponding actual x-coordinates. Similarly, the elements in the fourth row of the column matrix are multiplied one-to-one with each of the n elements (actual x-coordinates) in the fourth row of the transpose matrix (timestamps raised to the power of 3). These n products are then summed, and the result is equal to the sum of the products of the cubes of all timestamp values ​​and their corresponding actual x-coordinates. Through these calculations, a 4x1 column matrix is ​​obtained, completing the construction of the two components of the regular equation.

[0073] Solve the canonical equation to obtain the x-axis polynomial coefficients:

[0074] The regular equation takes the form of a square matrix × coefficient vector = column matrix, where the coefficient vector is a 4x1 matrix containing the four coefficients of the 3rd order polynomial on the x-axis (corresponding to the coefficients of the timestamp to the 0th, 1st, 2nd, and 3rd powers, respectively). The core of the solution is to obtain the coefficient vector through matrix operations.

[0075] Step 1: Determine if the square matrix is ​​invertible (calculate the determinant of the square matrix):

[0076] First, calculate the determinant of the 4x4 matrix (the determinant is a specific numerical value, which needs to be calculated using the expansion rules of a 4th-order determinant: expand the 4th-order determinant by a certain row or column, transforming it into a combination of four 3rd-order determinants, then calculate the value of each 3rd-order determinant separately, and finally superimpose them according to the expansion rules to obtain the result of the 4th-order determinant); if the determinant value is not equal to 0, it means that the matrix is ​​invertible, and the subsequent inversion operation can continue; if the determinant value is equal to 0, it means that the matrix is ​​not invertible, and the previous steps need to be checked again (such as whether there are any undeleted abnormal points in the cleaned trajectory points, whether the timestamps and coordinates of the coefficient matrix correspond incorrectly, whether the matrix transpose is correct, etc.), correct the problems, reconstruct the matrix and calculate the determinant, until the determinant value is not equal to 0.

[0077] Step 2, calculate the inverse matrix of the square matrix:

[0078] The inverse matrix is ​​a matrix whose product with the original square matrix is ​​an identity matrix (a 4x4 matrix with 1s on the main diagonal and 0s on the rest). The calculation process consists of three steps:

[0079] To calculate the cofactor matrix of a square matrix, for each element in the i-th row and j-th column (where i and j are integers from 1 to 4), first delete the i-th row and j-th column containing that element. The remaining elements form a smaller square matrix of 3 rows and 3 columns. Calculate the determinant of this smaller square matrix (i.e., the cofactor of that element). Then, based on the sign factor = (-1)... (i+j) The sign of i+j is positive when i+j is even and negative when i+j is odd. Multiply the cofactor by the sign factor to obtain the algebraic cofactor of the element. In this way, the algebraic cofactors of all 16 elements of the square matrix are calculated to form a 4x4 algebraic cofactor matrix.

[0080] Transpose the algebraic cofactor matrix:

[0081] Interchange the rows and columns of the cofactor matrix (i.e., the first row of the transpose matrix corresponds to the first column of the original cofactor matrix, the second row corresponds to the second column of the original matrix, and so on), to obtain the adjoint matrix; divide each element in the adjoint matrix by the previously calculated determinant value of the square matrix, and the new matrix obtained is the inverse matrix of the original square matrix (still 4 rows and 4 columns).

[0082] Step 3: Multiply the inverse matrix by the column matrix to obtain the coefficient vector:

[0083] Multiplying the 4x4 inverse matrix by the 4x1 column matrix yields a 4x1 coefficient vector. The four elements of this vector represent the coefficients of the 3rd order polynomial along the x-axis.

[0084] The elements in the first row of the coefficient vector correspond to the 0th power coefficients of the timestamps of the x-axis polynomial. They are calculated as follows: take the four elements of the first row of the inverse matrix, multiply them one-to-one with the elements of the first, second, third, and fourth rows of the column matrix, and then sum these four products. The elements in the second row of the coefficient vector correspond to the 1st power coefficients of the timestamps of the x-axis polynomial. They are calculated as follows: take the four elements of the second row of the inverse matrix, multiply them one-to-one with the elements of the first, second, third, and fourth rows of the column matrix, and then sum these four products. The elements in the third row of the coefficient vector... The timestamp coefficients of the x-axis polynomial are calculated as follows: take the four elements of the third row of the inverse matrix and multiply them one by one with the elements of the first, second, third, and fourth rows of the column matrix, then sum the four products. The elements of the fourth row of the coefficient vector correspond to the timestamp coefficients of the x-axis polynomial. They are calculated as follows: take the four elements of the fourth row of the inverse matrix and multiply them one by one with the elements of the first, second, third, and fourth rows of the column matrix, then sum the four products. Thus, the coefficients of the third-order x-axis polynomial are solved.

[0085] Verification of the accuracy of the x-axis polynomial coefficients:

[0086] Step 1, calculate the fitted x-axis coordinate for each timestamp:

[0087] Extract each timestamp from the set of cleaning trajectory points, and substitute them sequentially into the third-order polynomial of the x-axis composed of the solved coefficients to calculate the fitted x-axis coordinates corresponding to each timestamp; the substitution rule is as follows:

[0088] Fitted x-axis coordinates = (0th power coefficient × 1) + (1st power coefficient × timestamp value) + (2nd power coefficient × timestamp value squared) + (3rd power coefficient × timestamp value cubed); where the timestamp value squared is the timestamp value itself multiplied by itself, and the timestamp value cubed is the timestamp value squared and then multiplied by the timestamp value.

[0089] Step 2, calculate the deviation for each trajectory point:

[0090] The fitted x-axis coordinate corresponding to each timestamp is compared with the actual x-axis coordinate corresponding to that timestamp in the cleaning trajectory point set, and the deviation between the two is calculated: Deviation = |fitted x-axis coordinate - actual x-axis coordinate|. Taking the absolute value is to avoid positive and negative deviations canceling each other out and to truly reflect the degree of deviation.

[0091] Step 3: Calculate the average deviation and determine if the coefficient is acceptable.

[0092] Calculate the total deviation by adding the deviation values ​​of all cleaning trajectory points; calculate the average deviation by dividing the total deviation by the number of cleaning trajectory points n (i.e., average deviation = total deviation / n).

[0093] If the average deviation is ≤0.1 meters, it means that the x-axis polynomial coefficients are accurate and meet the trajectory fitting accuracy requirements. If the average deviation is >0.1 meters, all previous steps need to be checked again (such as whether the matrix transpose is wrong, whether the matrix multiplication element calculation is missing, whether the inverse matrix solution is wrong, whether there are any abnormalities in the cleaned trajectory points, etc.). After correcting the problems, the regular equation is reconstructed, the coefficients are solved and verified until the average deviation is ≤0.1 meters.

[0094] Solving and verifying the coefficients of the y-axis canonical equation; the solution process for the y-axis is completely consistent with that for the x-axis:

[0095] Construct an n x 4 matrix of coefficients for the y-axis (columns corresponding to the 0th, 1st, 2nd, and 3rd powers of the timestamps) and an n x 1 matrix of constant terms (columns corresponding to the actual y-axis coordinates of each cleaning trajectory point). After transposing the y-axis coefficient matrix, multiply it by the original coefficient matrix and the constant term matrix respectively to obtain a 4 x 4 matrix and a 4 x 1 matrix of columns for the y-axis, forming the y-axis regular equation. Calculate the determinant (to determine invertibility) and inverse matrix of the y-axis matrix, and then multiply the inverse matrix by the column matrix to obtain the four coefficients of the 3rd order polynomial of the y-axis (corresponding to the 0th to 3rd powers of the timestamps). Verify the y-axis coefficients by substituting the timestamps into the fitted y-axis coordinates and comparing them with the actual y-axis coordinates to calculate the average deviation. If the deviation is ≤0.1 meters, the coefficients are acceptable; otherwise, recalculate until the accuracy requirements are met.

[0096] The verified x-axis polynomial and y-axis polynomial are combined to form a two-dimensional model of the vehicle's actual driving trajectory. This model can calculate the corresponding lateral coordinates using the x-axis polynomial and the corresponding longitudinal coordinates using the y-axis polynomial based on any timestamp (within the time range of the clean trajectory point set). The two are combined to obtain the vehicle's fitted position at that timestamp. By concatenating the fitted positions of consecutive timestamps, the continuous driving trajectory of the vehicle within the controlled area can be accurately characterized, providing a reliable fitted trajectory benchmark for subsequent steps S203 to compare the trajectory with the standard path and calculate the deviation distance.

[0097] Step S203: Based on the actual vehicle trajectory model, perform a spatial projection comparison with a preset standard driving path to calculate the deviation distance between the two trajectories, specifically including:

[0098] First, retrieve the preset standard driving route data for the controlled area (such as roads around the campus) from the database. This data will record the standard route information that vehicles are allowed to travel, such as the specific location of the lane center line from the campus gate to the main road, the fixed trajectory range that vehicles should take (including the left and right boundary positions), and the road type corresponding to this standard route, whether it is a straight road, a curve, or an intersection. In addition, the data will also contain the coordinate system information used when this standard route was initially collected, such as which fixed object (such as a sculpture at the campus gate or a street lamp) was used as the origin of the coordinates, and which direction the x-axis and y-axis point (e.g., x-axis pointing east, y-axis pointing north), with the coordinate unit being meters. Then, organize this data into a time-coordinate corresponding format. If the original data only has spatial coordinates and no time information, assign a virtual timestamp to each coordinate point according to the average driving speed of vehicles in the controlled area to ensure that it can correspond to the time dimension of the actual trajectory later.

[0099] Since the actual driving trajectory model (obtained from step S202) and the preset standard path may use different coordinate systems—for example, the actual trajectory is based on the coordinate system of the onboard sensors, while the standard path is based on the coordinate system of the regional map—they are converted into the same coordinate system. The specific steps are as follows:

[0100] First, find three or more immovable landmarks within the controlled area (such as fixed traffic signs or roadside utility poles). Record the coordinates of these landmarks in the actual trajectory coordinate system and their coordinates in the standard path coordinate system. Then, based on the differences between the two sets of coordinates for these landmarks, calculate the parameters required for coordinate system transformation through spatial geometry calculations. This includes how much the origin needs to be offset (translation) and how much the coordinate axes need to be rotated (rotation angle). If the units of the two coordinate systems are different, the scaling ratio also needs to be calculated. Next, use these transformation parameters to transform all the coordinate points of the standard path into coordinates in the actual trajectory coordinate system. Finally, find one or two landmarks that were not used to calculate the transformation parameters and check the deviation between the coordinates of these landmarks in the standard path coordinate system and their coordinates in the actual trajectory coordinate system after transformation. If the deviation does not exceed 0.05 meters, the transformation is successful. If the deviation exceeds the limit, recalculate the transformation parameters and perform the transformation again.

[0101] The converted standard path and the actual driving trajectory model are projected onto the same two-dimensional plane (since most roads around the campus are flat, the impact of height differences does not need to be considered) to achieve spatial alignment. First, it is necessary to ensure that the time range matches: for example, if the timestamp of the actual trajectory is from 7:00 AM to 7:05 AM, then the portion of the standard path corresponding to this time period should be selected to ensure that the two correspond in time. If there is height data in the actual trajectory or the standard path (such as the vehicle passing through a small uphill), these coordinate points with height are converted into coordinates on the two-dimensional plane through vertical projection, so that the horizontal distance calculation will not be affected by the difference in height.

[0102] For each timestamp in the actual trajectory model, calculate the vehicle's position coordinates and their deviation from the standard path. First, find the point on the standard path closest to the actual position: this can be done by going through all the coordinate points on the standard path, or by generating more dense coordinate points on the standard path through piecewise interpolation. Then, calculate the straight-line distance between each standard point and the actual position, and the standard point with the smallest distance is the closest point. Next, use the straight-line distance between the actual position and this closest point as the trajectory deviation distance at that time point. If the actual position exceeds the allowable range of the standard path (e.g., driving into the oncoming lane), an additional penalty distance is added when calculating this deviation distance. For example, if the actual position exceeds the boundary by 1 meter, an additional 1-meter deviation distance is calculated, which can more clearly reflect the risk.

[0103] Step S204: Based on the deviation distance, a weighted normalization process is performed on the vehicle motion state parameters to obtain the vehicle trajectory deviation index, specifically including:

[0104] From the dynamic operation parameter dataset (output in step S106), find the vehicle motion parameters that correspond one-to-one with each timestamp of the actual trajectory. There are two main parameters: one is the real-time vehicle speed, which is the vehicle speed at each time point. It is necessary to ensure that there are no missing speed data. If the speed at a certain time point is not recorded, the average of the speeds before and after that time point is used to make up for it. At the same time, obviously unreasonable speeds should be removed (for example, a speed exceeding 36 km / h in a campus scene is an outlier). The other parameter is the real-time acceleration, which is the acceleration or deceleration of the vehicle at each time point (acceleration is a positive value, and deceleration is a negative value). Similarly, the data integrity should be checked and outliers should be removed (for example, the absolute value of acceleration exceeding 2 is rare in a campus and is an outlier).

[0105] Based on the safety requirements of the controlled area (e.g., a higher risk of trajectory deviation due to heavy pedestrian traffic, high vehicle speeds, or sudden acceleration / deceleration on campus), weighting rules for vehicle speed and acceleration are established to amplify deviation distances under high-risk conditions. Specifically:

[0106] The vehicle speed weight is divided into three levels. If the vehicle speed does not exceed 10.8 km / h, it is considered low-speed driving within the campus, and the weight is set to 1.0, which is the basic weight. If the vehicle speed is between 10.8 and 21.6 km / h, it is considered normal speed around the campus, and the weight is set to 1.5, which is the medium weight. If the vehicle speed exceeds 21.6 km / h, it exceeds the speed limit around the campus, and the weight is set to 2.0, which is the high weight.

[0107] Acceleration weights are also divided into three levels. If the absolute value of acceleration does not exceed 0.5, the vehicle is driving smoothly without sudden acceleration or deceleration, and the weight is set to 1.0, which is the basic weight. If the absolute value is between 0.5 and 1.5, it is considered a slight sudden acceleration or deceleration, and the weight is set to 1.5, which is the medium weight. If the absolute value exceeds 1.5, it is considered a significant sudden acceleration or deceleration, which is likely to cause danger, and the weight is set to 2.0, which is the high weight.

[0108] Finally, the overall weight is calculated by multiplying the speed weight and acceleration weight at each time point to obtain the overall weight at that time point. For example, if the speed weight is 1.5 and the acceleration weight is 2.0 at a certain time point, the overall weight is 3.0. In this way, the deviation distance under high-risk conditions will be multiplied by a larger number, which better reflects the risk.

[0109] For each time point, multiply the deviation distance calculated in step S203 by the comprehensive weight of that time point to obtain the weighted deviation value; for example, if the deviation distance at a certain time point is 0.5 meters and the comprehensive weight is 3.0, then the weighted deviation value is 0.5 × 3.0 = 1.5; after the calculation, the result should be checked. If the weighted deviation value is 0, it means that there is no deviation at this time point, or if a negative number appears, then go back to check the original data or calculation process to ensure that all weighted deviation values ​​are non-negative positive numbers.

[0110] The weighted deviation values ​​at all time points are converted into an index within a fixed range of 0-1, ensuring consistent comparison regardless of the original values. First, the maximum and minimum values ​​of all weighted deviation values ​​are identified (typically, the minimum is 0, representing a time point with no deviation). Then, for each weighted deviation value, the normalized deviation value is calculated by subtracting the minimum value from the current value and dividing by the result of subtracting the minimum from the maximum value. If all weighted deviation values ​​are the same (maximum equals minimum), all normalized deviation values ​​are set to 0.5. Finally, all normalized deviation values ​​are summed and divided by the number of time points to obtain the vehicle trajectory deviation index. The closer this index is to 1, the higher the overall risk of trajectory deviation; the closer it is to 0, the closer the vehicle is to the standard path, and the lower the risk.

[0111] In a preferred embodiment of the present invention, step S3, which involves assessing the collision risk based on the vehicle trajectory deviation and the dynamic operating parameters, and calculating the minimum safe distance between the vehicle and the boundary, includes:

[0112] Step S301: Based on the vehicle trajectory deviation index and the dynamic operation parameter dataset, extract the vehicle's current position coordinates, speed, heading angle parameters, and geometric information of the control area boundary. Specifically, this includes: determining the data source and association logic; using the vehicle's unique identifier ID as a link, associating the vehicle trajectory deviation index output in step S204 with the dynamic operation parameter dataset generated in step S106 to ensure that all data correspond to the same target vehicle and avoid confusion between data from different vehicles.

[0113] Next, extract the vehicle's current core motion parameters. From the dynamic operation parameter dataset, select the latest data with the latest timestamp to obtain the vehicle's current position coordinates, i.e., the x-axis horizontal coordinate and y-axis vertical coordinate corresponding to that timestamp. At the same time, extract the current speed, and confirm that the speed is instantaneous and without outliers (e.g., in a campus scenario, if the speed exceeds 36 km / h, check whether it is a data error before extracting). In addition, extract the current heading angle parameter, which must be consistent with the vehicle's real-time attitude. For example, when the vehicle is traveling due east, the heading angle should match the eastward reference direction of the coordinate system.

[0114] Finally, the geometric information of the control area boundary is obtained. The boundary data of the area is retrieved from the preset control area geographic database, including the type of boundary (such as guardrails around the campus, the edge of the sidewalk, the boundary of the green belt, the road isolation pier, etc.), the spatial distribution of the boundary, and the characteristic line segment information after each boundary is broken down. Each characteristic line segment must include the coordinates of the two endpoints (to determine the x-axis and y-axis values ​​of the start and end points of the line segment in the same coordinate system), and the attributes of the line segment (such as whether it is a rigid boundary, whether it belongs to the pedestrian protection zone boundary, etc.) are marked to ensure that the boundary information is complete and consistent with the coordinate system of the vehicle's current position coordinates.

[0115] Step S302: Based on the vehicle's current position and heading angle parameters, construct a predicted motion spatial geometry with the vehicle position as the origin and the heading direction as the main axis. Specifically, this includes setting the vehicle's current position coordinates extracted in step S301 as the origin of the predicted motion spatial geometry. This origin is the reference point for all subsequent spatial position calculations, ensuring that all predicted positions are calculated around the vehicle's current position.

[0116] Then, based on the heading angle parameters extracted in step S301, the current driving direction of the vehicle is set as the main axis direction of the geometric structure. For example, if the heading angle shows that the vehicle is currently driving due north, the main axis direction is completely consistent with the due north direction in the coordinate system. If the heading angle is in the northeast direction, the main axis direction is simultaneously adjusted to the northeast direction to ensure that the main axis can accurately reflect the actual driving trend of the vehicle. Next, combining the vehicle's current speed and acceleration parameters (extracted from the dynamic operating parameter dataset), the positions the vehicle may reach at multiple future time points (e.g., 0.5 seconds, 1 second, 1.5 seconds) are predicted. For example, if the vehicle's current speed is 10 km / h and it is traveling at a constant speed, the distance the vehicle may travel along the main axis in the next second can be calculated, and the predicted position at that time point can be determined. If there is acceleration, the distance of the predicted position needs to be adjusted according to the acceleration or deceleration trend (the predicted position is farther when accelerating and closer when decelerating). These predicted positions at different time points are connected in chronological order to form a polygonal spatial region centered on the main axis. This region is the spatial range in which the vehicle may move in the future, which is the predicted motion spatial geometry. Finally, it is checked whether this geometry covers the possible extreme values ​​of the vehicle's motion, such as the extreme positions during rapid acceleration or deceleration, to ensure that the structure can fully reflect the future motion trend of the vehicle.

[0117] Step S303: Traverse all feature line segments of the control area boundary, and calculate the perpendicular foot of the vehicle's current position point to the line containing each boundary line segment; determine whether the perpendicular foot is within the effective interval of the corresponding boundary line segment. If it is within the effective interval, record the Euclidean distance from the perpendicular foot to the vehicle's position point as a candidate distance; if the perpendicular foot is outside the effective interval, calculate the Euclidean distance from the vehicle's position point to the two endpoints of the boundary line segment and record the minimum value as a candidate distance; select the minimum spatial distance from the candidate distances corresponding to all boundary line segments; based on the vehicle's current speed and acceleration parameters, dynamically compensate and correct the minimum spatial distance to obtain dynamic distance parameters, specifically including:

[0118] In the boundary geometry information obtained in step S301, each boundary is decomposed into continuous feature segments (e.g., a 100-meter guardrail can be divided into 10 10-meter feature segments, each with definite start and end coordinates), and a unique identifier is assigned to each feature segment to avoid duplicate calculations or omissions. Then, starting from the feature segments near the vehicle's current position (e.g., taking the vehicle's current position as the center, priority is given to calculating boundary segments within a 50-meter range), the calculation is gradually expanded to all feature segments in the entire controlled area. This improves calculation efficiency and avoids calculating segments that are too far away and have minimal impact on the current risk assessment first.

[0119] For each feature line segment being traversed, first determine the straight line containing the segment (based on the coordinates of the two endpoints of the segment, determine the spatial position of the straight line), then calculate the foot of the perpendicular from the vehicle's current position to this straight line (i.e., the intersection of the perpendicular line drawn from the current position to the straight line); next, determine whether the foot of the perpendicular is within the valid interval of the feature line segment. The valid interval refers to the area between the two endpoints of the feature line segment, determined by checking whether the coordinates of the foot of the perpendicular are within the range of the coordinates of the two endpoints; for example, if the starting coordinates of the segment are (x1, y1) and the ending coordinates are (x2, y2), then the foot of the perpendicular needs to be confirmed. The x-axis coordinate of the perpendicular must be between x1 and x2 (if x1 < x2, then the x-coordinate of the perpendicular foot must satisfy x1 ≤ perpendicular foot x ≤ x2; if x1 > x2, then x2 ≤ perpendicular foot x ≤ x1). Simultaneously, the y-axis coordinate of the perpendicular foot must be between y1 and y2 (the judgment logic is consistent with the x-axis). If the perpendicular foot is within the valid interval, the straight-line distance from the perpendicular foot to the vehicle's current position is directly recorded as the candidate distance corresponding to the feature line segment. If the perpendicular foot is outside the valid interval, the straight-line distances from the vehicle's current position to the start and end points of the feature line segment are calculated respectively, and the smaller of these two distances is taken as the candidate distance corresponding to the feature line segment.

[0120] After calculating the candidate distances for all feature line segments, the distance with the smallest value is selected from all candidate distances. This distance is the minimum spatial distance between the vehicle's current position and the boundary of the controlled area. Then, based on the vehicle's current speed and acceleration parameters, the minimum spatial distance is dynamically compensated and corrected: if the vehicle's current speed is high (e.g., exceeding 21.6 km / h in a campus scenario), the compensation amount needs to be increased because the braking distance required when the vehicle is traveling at high speed is longer, and the actual safe distance needs to be greater than the statically calculated minimum spatial distance; if the vehicle is accelerating rapidly (the acceleration is positive and the absolute value is large), the compensation amount also needs to be increased, as it is predicted that the vehicle may approach the boundary faster; if the vehicle is decelerating rapidly (the acceleration is negative and the absolute value is large), the compensation amount can be appropriately reduced because the risk of approaching the boundary is reduced as the vehicle is decelerating. The distance after compensation and correction is the dynamic distance parameter, which can better match the safety distance requirements under the actual movement state of the vehicle.

[0121] Step S304: Based on the dynamic distance parameter, a weighted fusion process is performed using the vehicle trajectory deviation index to obtain the minimum safe distance between the vehicle and the boundary. Specifically, this includes: considering the safety requirements of the controlled area (e.g., in densely populated campus areas, the risk of trajectory deviation has a greater impact on safety), setting the weights of the dynamic distance parameter and the vehicle trajectory deviation index; generally, the vehicle trajectory deviation index has a higher weight (e.g., a weight coefficient of 0.6), while the dynamic distance parameter has a lower weight (e.g., a weight coefficient of 0.4); if the vehicle's current speed exceeds the speed limit, the weight of the trajectory deviation index can be appropriately increased (e.g., adjusted to 0.7) to further strengthen the impact of deviation risk on the safe distance.

[0122] Then, the dynamic distance parameter obtained in step S303 is multiplied by its corresponding weight coefficient, and the vehicle trajectory deviation index obtained in step S204 is multiplied by its corresponding weight coefficient. The two product results are added together to obtain the fused value. Finally, it is checked whether the fused value meets the safety standards of the controlled area (e.g., the minimum safe distance in the campus area shall not be less than 1 meter). If the value is lower than the safety standard, it needs to be adjusted according to the lower limit of the safety standard. If the value meets the standard, then the value is the minimum safe distance between the vehicle and the boundary. This distance takes into account the actual distance between the vehicle and the boundary and the risk of trajectory deviation, and can more accurately reflect the collision risk faced by the vehicle.

[0123] In a preferred embodiment of the present invention, step S4 involves performing spatial modeling analysis on the dynamically selected vehicle trajectory curvature center point, vehicle safety envelope boundary point, and historical trajectory weighted center point based on the minimum safety distance, to obtain a geometric feature model, including:

[0124] Step S401: Based on the minimum safe distance between the vehicle and the boundary, extract the three-dimensional coordinate data of the curvature center point of the vehicle's motion trajectory, the spatial distribution information of the vehicle's safety envelope boundary points, and the spatiotemporal feature parameters of the weighted center point of the historical trajectory. Then, normalize the corresponding spatial coordinate data to obtain the normalized coordinates of the curvature center point of the vehicle's motion trajectory, the normalized coordinates of the vehicle's safety envelope boundary points, and the normalized coordinates of the weighted center point of the historical trajectory. Specifically, this includes: using the minimum safe distance between the vehicle and the boundary obtained in step S3 as the core reference, defining the effective range of data extraction, and extracting only the relevant point data of the vehicle within the risk-related area corresponding to the minimum safe distance, avoiding the inclusion of redundant information from irrelevant areas, and ensuring that the data is directly related to the collision risk assessment.

[0125] Next, key data for the three types of center points are extracted respectively:

[0126] Extract the three-dimensional coordinate data of the center point of the vehicle's motion trajectory curvature. From the actual driving trajectory model of the vehicle constructed in step S2, select road segments with curved trajectories (such as turning and lane-changing areas), calculate the trajectory curvature of each curved road segment (the greater the curvature, the more obvious the trajectory curvature), and determine the center point corresponding to the curvature (i.e., the center of the curved arc of the trajectory). Record the three-dimensional coordinates of the center point (including the x-axis horizontal, y-axis vertical, and z-axis height, the z-axis height is determined according to the actual height of the road surface, such as a fixed value for the z-axis coordinate on a flat road surface).

[0127] Extract the spatial distribution information of the vehicle safety envelope boundary points: Combine the vehicle static feature data (vehicle model, body size) obtained in step S1, and take the current position of the vehicle as the center, determine the boundary of the vehicle safety envelope according to the body size and the minimum safe distance range. For example, if the width of a small car body is 1.8 meters and the minimum safe distance is 1 meter, then the left and right boundaries of the safety envelope are extended by 1 meter to both sides of the body, forming an envelope range with a width of 3.8 meters. Within this range, select multiple boundary points evenly (such as the left and right sides of the front and rear ends of the body, and the left and right sides of the middle of the body), and record the spatial coordinates and distribution position of each boundary point, such as the left boundary point of the front end of the body and the right boundary point of the middle of the body.

[0128] Extract the spatiotemporal feature parameters of the historical trajectory weighted center point. From the vehicle spatiotemporal trajectory information generated in step S104, select historical trajectory points within the most recent period (e.g., the past 30 seconds). According to the time weight allocation rule (the more recent the trajectory point, the higher the weight; for example, the weight of trajectory points within the most recent 5 seconds is set to 0.8, those within 5-15 seconds to 0.5, and those within 15-30 seconds to 0.2), calculate the weighted average of the coordinates of each historical trajectory point to obtain the spatial coordinates of the historical trajectory weighted center point. At the same time, extract the spatiotemporal feature parameters of the center point, including the corresponding timestamp range (e.g., the start and end times of the past 30 seconds) and the average speed of the vehicle within this time period.

[0129] Finally, the spatial coordinates of the three types of center points are normalized, uniformly mapping all coordinate data to a fixed interval of 0-1, eliminating the impact of numerical differences in different coordinate dimensions (such as x-axis range of 0-100 meters and y-axis range of 0-80 meters) on subsequent modeling. For example, if the original x-axis coordinate range is 0-100 meters and the x-axis coordinate of a certain curvature center point is 50 meters, then the normalized x-axis coordinate is 0.5. Similarly, the y-axis and z-axis coordinates are processed to finally obtain the normalized coordinates of the vehicle motion trajectory curvature center point, the normalized coordinates of the vehicle safety envelope boundary point, and the normalized coordinates of the historical trajectory weighted center point.

[0130] Step S402: Based on the normalized coordinates of the weighted center points of the historical trajectory, an initial surface mesh is constructed. The coordinates of the curvature center points of the normalized vehicle motion trajectory are applied as geometric constraints to the initial surface mesh, and the coordinates of the boundary points of the vehicle safety envelope are used as boundary conditions to obtain a spatial geometric feature surface representing the spatial distribution characteristics of vehicle motion. Specifically, this includes: constructing an initial surface mesh based on the coordinates of the weighted center points of the normalized historical trajectory; arranging the weighted center points of the historical trajectory according to time sequence or spatial distribution rules; and constructing basic mesh units (such as triangular meshes or quadrilateral meshes) with adjacent center points as vertices. For example, if there are 10 weighted center points of the historical trajectory, they are connected sequentially according to time sequence, and then the center points of adjacent time points are connected by horizontal line segments to form a grid-like initial surface mesh. This mesh needs to cover the core area of ​​the vehicle's historical motion to ensure that it can reflect the past motion space range of the vehicle.

[0131] Next, surface geometric constraints are applied, using the coordinates of the normalized vehicle trajectory curvature center point as constraints to adjust the shape of the initial surface mesh. For areas with significant trajectory curvature (corresponding to the curvature center point), the mesh cell density is increased, such as adjusting the original 1-meter interval mesh to a 0.5-meter interval. Simultaneously, the mesh is made to conform to the curvature trend of the trajectory near the curvature center point (e.g., when turning left, the left side of the mesh shifts to the left with the curvature center), ensuring the surface accurately reflects the curvature characteristics of the vehicle trajectory. Then, surface boundary constraints are applied, using the coordinates of the normalized vehicle safety envelope boundary points as the surface boundary limits. The edges of the initial surface mesh are trimmed. If the initial mesh exceeds the area enclosed by the safety envelope boundary points, the excess mesh cells are deleted; if the mesh does not cover the complete safety envelope, mesh cells near the boundary are added, ensuring the final surface boundary is completely consistent with the safety envelope boundary, preventing the surface from exceeding the vehicle's actual possible movement space.

[0132] Finally, the spatial geometric feature surface is obtained by integrating the constraints. After adjusting the geometric constraints and boundary constraints, the initial surface mesh forms a surface that can completely represent the spatial distribution of vehicle motion. This surface includes the trajectory trend of historical motion (from the weighted center point of historical trajectory), reflects the curvature characteristics of the current trajectory (from the curvature center point), and also defines the spatial boundary of motion (from the safety envelope boundary point).

[0133] Step S403: Based on the spatial geometric feature surface, calculate the corresponding rate of curvature change and gradient direction to obtain a parameterized model containing spatial geometric features. Specifically, this includes: uniformly selecting multiple sampling points on the spatial geometric feature surface (e.g., selecting one point every 0.2 meters); for each sampling point, selecting 3-5 adjacent points around it to form a small curve segment; calculating the curvature of the curve segment (the larger the curvature value, the more significant the curvature of the curve segment); then calculating the curvature difference between adjacent sampling points, and dividing it by the spatial distance between adjacent sampling points to obtain the rate of curvature change. The rate of curvature change can reflect the changing trend of the curvature of the surface. For example, if the curvature of a road segment gradually increases from 0.1 to 0.5, and the rate of curvature change is positive, it indicates that the curvature of the trajectory in this area is gradually increasing. Next, for each sampling point on the surface, the slope (i.e., the rate of change of height of the point along a certain axis) in the x-axis, y-axis, and z-axis directions is calculated, and the direction with the largest slope is identified as the gradient direction of that point. The gradient direction can reflect the tilt trend of the surface. For example, if the slope of the surface is the largest along the positive x-axis at a certain sampling point, it indicates that the point is tilted in the positive x-axis direction, which can indirectly reflect the possible deviation trend of the vehicle's motion direction in that area. Finally, the rate of change of curvature, gradient direction, and normalized coordinates of each sampling point are used as core parameters and organized into structured data according to the correspondence between the sampling point coordinates, rate of change of curvature, and gradient direction.

[0134] Step S404 involves feature extraction and dimensionality reduction of the parameterized model to form a geometric feature model representing the spatial characteristics of vehicle motion. Specifically, this includes: selecting parameters from the core parameters of the parameterized model that characterize key spatial features of vehicle motion; for example, extracting sampling points with the largest rate of curvature change (corresponding to areas where the vehicle trajectory curves most sharply, potentially high-risk turning points), areas with the most significant gradient direction changes (corresponding to areas where the vehicle's motion direction may abruptly change), and sampling points near the safety envelope boundary (corresponding to risk areas where the vehicle approaches the boundary); simultaneously, retaining parameters corresponding to the weighted center points of historical trajectories (reflecting the core trends of past vehicle motion), and removing redundant parameters with low correlation to motion risk (such as repeated sampling point parameters in flat surface areas); then, since the extracted feature parameters may contain multiple dimensions (such as sampling point coordinates, rate of curvature change, gradient direction, etc.), it is necessary to reduce the number of parameters through data dimensionality reduction methods (such as retaining core dimensions and merging similar parameters), for example... By merging sampling point parameters with similar curvature change rates and adjacent spatial locations into a group, the average curvature change rate of this group is used to represent the overall characteristics; or dimensions with minimal impact on risk assessment are eliminated (e.g., in scenarios where the z-axis height change is small, z-axis-related parameters can be ignored); finally, the dimensionality-reduced feature parameters are logically correlated according to spatial location and key geometric features to construct the final geometric feature model; three types of core point data are extracted based on the minimum safe distance between the vehicle and the boundary, which can accurately screen out key spatial information directly related to collision risk, avoid including redundant data in irrelevant areas, and reduce the interference of invalid information on modeling; by calculating the curvature change rate of the spatial geometric feature surface (reflecting the dynamic change of the trajectory curvature, such as the transition from a gentle curve to a sharp curve) and gradient direction (reflecting the surface tilt trend, indirectly reflecting the possible deviation of the vehicle's motion direction), the abstract surface shape is transformed into quantifiable specific parameters, solving the problem that traditional geometric models are difficult to accurately describe motion spatial characteristics and cannot provide definite indicators for subsequent analysis.

[0135] In a preferred embodiment of the present invention, the geometric feature model is converted into a corresponding virtual feature point set, and a dynamic evaluation structure is constructed based on the virtual feature point set, including:

[0136] Step S405: Based on the geometric feature model, perform feature space mapping to convert high-dimensional geometric features into low-dimensional feature vectors. Specifically, this includes: extracting all parameters characterizing the spatial features of vehicle motion from the geometric feature model formed in step S404, including but not limited to the curvature change rate of the spatial geometric feature surface (curvature change values ​​at different sampling points), gradient direction (tilt trend parameters of each sampling point in the x-axis, y-axis, and z-axis directions), the coordinates of the normalized vehicle motion trajectory curvature center point, the coordinates of the normalized vehicle safety envelope boundary point, and the coordinates of the weighted center point of the historical trajectory, and calculating the total dimension corresponding to these parameters (for example, the curvature change rate corresponds to 50 sampling point parameters, the gradient direction corresponds to 50 sampling point parameters, and the three types of center point coordinates each correspond to 3 dimensions, with a total dimension that may reach more than 110 dimensions), and determining the composition of the high-dimensional features.

[0137] Next, based on the needs of vehicle collision detection, and using the principle of strong correlation with collision risk as the selection criterion, feature dimensions that have a direct impact on risk assessment are prioritized for retention, while redundant or weakly correlated dimensions are eliminated (for example, the z-axis height parameter has a minimal impact on collision risk in flat controlled areas (such as campus roads) and can be treated as a secondary dimension for weakening). At the same time, referring to the minimum safe distance between the vehicle and the boundary obtained in step S3, feature dimensions corresponding to areas close to the boundary and with smaller minimum safe distances are set as key retention objects in the mapping process to ensure that the spatial characteristics of high-risk areas can still be reflected after mapping.

[0138] Then, a mapping method adapted to spatial features is used to gradually convert high-dimensional geometric features into low-dimensional feature vectors. For example, sampling point parameters with similar curvature change rates and adjacent spatial locations are merged into a curvature trend feature term, and the average curvature change rate of the region is used to represent the overall trend. Safety envelope boundary point parameters with consistent gradient directions are integrated into a boundary tilt feature term to simplify the complexity of the direction dimension. At the same time, the coordinates of the three types of center points are integrated according to the priority of distance from the boundary, and the center point coordinate features corresponding to the minimum safety distance are retained first. Through this process, the high-dimensional features of more than 100 dimensions are compressed into low-dimensional feature vectors of 10-20 dimensions, and each vector dimension corresponds to a specific vehicle motion spatial feature (such as near-boundary curvature trend, safety envelope tilt direction, and historical trajectory center offset).

[0139] Finally, the mapped low-dimensional feature vectors are correlated and verified with the original high-dimensional geometric features to check whether the low-dimensional vectors completely retain the key information related to collision risk in the original model (such as whether the feature of a sudden increase in curvature in a certain region in the original model can still be reflected by the curvature trend feature term in the low-dimensional vector). If there is a lack of key information, the mapping rules are readjusted (such as increasing the feature dimension weight corresponding to this type of information) until the low-dimensional feature vectors can accurately reflect the core risk correlation attributes of the original high-dimensional geometric features.

[0140] Step S406 involves performing cluster analysis on the low-dimensional feature vectors to obtain a representative set of virtual feature points. This set of virtual feature points includes key spatial location information and feature weight parameters. Specifically, this includes: performing consistency adjustment on the low-dimensional feature vectors obtained in step S405 to ensure that all parameters of each dimension of all vectors are within the same data range (e.g., coordinate parameters that have been normalized are kept in the 0-1 range, and parameters such as the rate of curvature change are mapped to the 0-1 range through linear adjustment), avoiding clustering bias caused by differences in the scale of data of different dimensions; and removing abnormal vectors (e.g., vectors whose rate of curvature change is far beyond the physically reasonable range due to temporary sensor interference), ensuring that all vectors participating in the clustering can truly reflect the spatial characteristics of vehicle motion.

[0141] Next, the rules and objectives of clustering analysis are determined: in conjunction with the vehicle collision detection scenario, spatial distribution similarity and risk level similarity are used as dual clustering criteria. Whether the spatial distribution similarity index corresponds to adjacent spatial locations (e.g., feature vectors in the same turning area are grouped into one class), and whether the collision risks corresponding to the risk level similarity index are similar (e.g., vectors with small minimum safe distance and large rate of curvature change are grouped into high-risk classes). The clustering objective is set to cover all key risk areas without redundancy. The number of clusters is usually controlled to 5-10 (e.g., divided into near-boundary high-risk classes, normal trajectory classes, and historical offset classes, etc.), which avoids both too many categories leading to complex subsequent evaluation and too few categories causing omission of key features.

[0142] Then, a clustering method adapted to spatial features is used to group the preprocessed low-dimensional feature vectors. For example, vectors that are spatially adjacent and all have a high-risk level are clustered into one class. The average value of each dimension parameter of all vectors in this class is calculated to obtain a class center vector. Then, based on the spatial location information corresponding to the class center vector (such as normalized coordinates extracted from the vector), a representative point of this class is determined, which is a virtual feature point. In this way, a virtual feature point is generated for each class of vectors, ultimately forming a set containing 5-10 virtual feature points. Finally, a corresponding weight is assigned according to the degree of risk association of each virtual feature point to its class. For example, virtual feature points in the near-boundary high-risk class are assigned a weight of 0.8-1.0 because they are directly associated with collision risk; virtual feature points in the normal trajectory class are assigned a weight of 0.2-0.3 because they have a lower risk; and virtual feature points in the historical offset class are assigned a weight of 0.5-0.7 because they reflect the trajectory deviation trend. The weight allocation should refer to the minimum safe distance in step S3 (the smaller the distance, the higher the weight) and the trajectory deviation in step S2 (the higher the deviation, the higher the weight) to ensure that the weight can accurately reflect the risk level corresponding to the virtual feature point, and finally form a set of virtual feature points containing key spatial location information (normalized coordinates) and feature weight parameters.

[0143] Step S407: Based on the spatial distribution characteristics and weight parameters of the virtual feature point set, a multi-level dynamic evaluation structure is constructed. The multi-level dynamic evaluation structure includes a spatial relationship layer, a feature weight layer, and a risk assessment layer. Specifically, it includes: calculating the spatial distance (e.g., Euclidean distance) between any two virtual feature points based on the key spatial location information (normalized coordinates) of each virtual feature point, and determining their relative positional relationship (e.g., point A is 0.3 units northeast of point B, and point C is close to the boundary of the control area fence); at the same time, the positional relationship between each virtual feature point and the boundary of the control area is included in the analysis (e.g., the normalized distance between point D and the fence boundary is 0.1 and point E is far away from all boundaries), and determining the spatial distribution pattern of each point within the control area.

[0144] Then, using virtual feature points as nodes, and with spatial distances ≤0.2 units (corresponding to actual distances converted according to normalization ratios) or belonging to the same high-risk area as connection conditions, related nodes are connected with line segments to form a spatial relationship network. For example, three virtual feature points in the near-boundary high-risk category are connected to form a high-risk node cluster because they are spatially adjacent and all are high-risk; two points in the normal trajectory category are distributed on normal driving paths, maintaining a certain distance from the high-risk node cluster and not directly connected. This network can intuitively present the spatial relationships of each virtual feature point, providing a spatial dimension basis for subsequent risk assessment.

[0145] The feature weight parameters of each virtual feature point obtained in step S406 are entered into the feature weight layer as basic weight values ​​to form a correspondence table between virtual feature points and basic weights, ensuring that the initial weight of each point is searchable and traceable.

[0146] Then, the base weights are adjusted by combining the vehicle's real-time motion status (extracting the latest vehicle speed and acceleration from the dynamic operation parameter dataset in step S1) and environmental changes (such as whether there are temporary pedestrians crossing the controlled area or whether the weather is severe). For example, if the vehicle's current speed increases by 50% compared to before, the weights of all virtual feature points are multiplied by a coefficient of 1.2 (risk is amplified when driving at high speed); if a pedestrian suddenly appears in the area corresponding to a virtual feature point of high risk near the boundary, the weight of that point is increased by an additional 0.3; if the weather is clear and there is no interference, the weights remain unchanged at the base value. The adjusted weights need to be updated in real time to ensure that they reflect the actual risks in the current scenario.

[0147] Finally, based on the spatial relationship layer network, a weight transfer logic is set. The risk impact of high-weight nodes (e.g., weight ≥ 0.8) can be transferred to adjacent nodes, with a transfer ratio of 30% of the high-weight node's weight (e.g., high-weight node A has a weight of 0.9, and adjacent node B has a weight of 0.5; after the transfer, node B's weight becomes 0.5 + 0.9 × 30% = 0.77). Low-weight nodes (e.g., weight ≤ 0.3) do not transfer their weights to other nodes to avoid low-risk interference with high-risk assessments. Through weight transfer, the feature weight layer can reflect the spatial diffusion effect of risk, which is more in line with the actual propagation law of collision risk.

[0148] Construct a risk assessment layer:

[0149] First, risk assessment indicators and rules are established. Based on the node associations in the spatial relationship layer and the dynamic weights in the feature weight layer, core risk assessment indicators are determined, including the number of high-weight nodes, the spatial range of high-risk node clusters, and the total node weights. Simultaneously, assessment rules are formulated. For example, a high-risk association is defined as having ≥3 high-weight nodes or a total node weight ≥2.5; a low-risk association is defined as having 1 high-weight node and a total node weight <1.0. These rules should refer to the minimum safety distance (the smaller the distance, the lower the high-risk threshold) from step S3 and historical warning data (such as risk assessment standards in similar past scenarios). Then, the node connection information from the spatial relationship layer and the dynamic weight data from the feature weight layer are input into the assessment calculation framework, which will automatically perform statistical analysis. The system calculates the number of high-weight nodes, the total weight of all nodes, identifies the range of high-risk node clusters, and outputs initial risk assessment results according to the assessment rules (e.g., if the current scenario is medium to high risk, the high-risk node clusters are located near the guardrail boundary). Finally, because vehicles are in continuous motion, the spatial location and weight of virtual feature points change over time (e.g., when a vehicle approaches the boundary, the weight of high-risk nodes near the boundary increases). The risk assessment layer needs to receive these updated data in real time, re-execute the assessment calculation every 0.5 seconds, and update the risk assessment results. Simultaneously, if the environment of the controlled area changes (e.g., temporary roadblocks are set up), the assessment rules need to be adjusted accordingly (e.g., adding rules to increase the weight of nodes near roadblocks) to ensure that the entire multi-level dynamic assessment structure can continuously adapt to vehicle movement and environmental changes.

[0150] In a preferred embodiment of the present invention, step S5, based on the dynamic evaluation structure, calculates the security threat level to obtain a vehicle security threat index, including:

[0151] Step S501: Based on the multi-level dynamic evaluation structure, extract the spatial distribution features of the spatial relationship layer, the weight parameters of the feature weight layer, and the initial evaluation results of the risk assessment layer. Specifically, since the multi-level dynamic evaluation structure will be updated in real time with the vehicle movement, it is necessary to lock the current timestamp (to be consistent with the timestamp of the vehicle's current movement state data), and only extract the latest data of the dynamic evaluation structure under this timestamp to avoid using lagging or expired historical data and ensure that the extracted content matches the current actual state of the vehicle.

[0152] Next, extract key information hierarchically:

[0153] Extracting the spatial distribution features of the spatial relationship layer: From the spatial relationship layer, obtain the specific spatial location (normalized x-axis and y-axis coordinates) of virtual feature points, determine the relative distance between each point and the boundary of the controlled area (e.g., the normalized distance of a point from the fence boundary is 0.1); count the number of high-risk node clusters (e.g., a cluster consisting of 3 adjacent high-weight virtual feature points is 1 high-risk node cluster) and the spatial coverage of each cluster (e.g., a high-risk cluster covers 0.2-0.5 units in the x-axis direction and 0.3-0.6 units in the y-axis direction); record the connection relationship between virtual feature points (e.g., high-risk node A is directly connected to node B, and node B is not connected to low-risk node C), and mark the tightness of the connection (e.g., the smaller the spatial distance between adjacent nodes, the tighter the connection).

[0154] Extract the weight parameters of the feature weight layer: Obtain the complete weight information of each virtual feature point from the feature weight layer, including the initial basic weight (the weight assigned in step S406), the dynamically adjusted current weight (the weight adjusted by combining the real-time motion state and the environment), and the final weight after weight transfer (the final value after receiving weights transferred from other nodes); at the same time, extract detailed records of weight transfer, such as which node transferred weights to which node, the proportion of transfer (e.g., node A transferred 30% of the weights to node B), and the amount of weight change after transfer, to ensure that the weight parameters are traceable.

[0155] From the risk assessment layer, obtain the initial risk judgment conclusion (e.g., the current scenario is medium risk and there is one high-risk node cluster); extract the core indicator data supporting this judgment, including the number of high-weight nodes (weight ≥ 0.8), the total weight of all virtual feature points, and the spatial coverage ratio of the high-risk node cluster (e.g., the high-risk cluster covers 20% of the core driving area of ​​the control zone); at the same time, record the rule version used in the initial assessment (e.g., based on the risk assessment rule updated in May 2024) to facilitate subsequent verification of the rule applicability.

[0156] Finally, check for missing data (e.g., the current weight of a virtual feature point is not recorded, or the spatial range of a high-risk node cluster is not labeled). If there are missing data, return to the multi-level dynamic evaluation structure construction step (step S407) to complete the data. At the same time, verify the consistency of data in different layers. For example, the weight of the labeled high-risk node A in the spatial relationship layer must meet the high weight standard (≥0.8) in the feature weight layer. If there is inconsistency, the weight calculation process must be rechecked to ensure that the extracted data is accurate and complete and can support the subsequent calculation of security threat level.

[0157] Step S502 involves comprehensively calculating various characteristic parameters based on the vehicle trajectory deviation index and the minimum safe distance between the vehicle and the boundary, and dynamically adjusting the weighting coefficients of each characteristic parameter according to the real-time movement state of the vehicle and changes in the surrounding environment to obtain the vehicle safety threat index. Specifically, this includes:

[0158] The spatial distribution features (number of high-risk node clusters, coverage area of ​​high-risk clusters) extracted in step S501, the weight parameters (final weight sum of all virtual feature points, number of high-weight nodes), and the initial evaluation results (initial risk level) are integrated with the vehicle trajectory deviation index output in step S2 (e.g., 0.6, the closer to 1, the higher the deviation risk) and the minimum safe distance between the vehicle and the boundary output in step S3 (e.g., 1.2 meters, the smaller the distance, the higher the collision risk) into a unified set of basic parameters for comprehensive calculation, ensuring that all parameters are the latest data under the current timestamp, without data lag or misalignment.

[0159] Then, based on the safety requirements of the controlled area (such as the need for strict control of collision risks in campus areas), initial calculation coefficients are assigned to each basic parameter to establish the basic calculation logic. For example, the basic score is increased by 20 for each additional high-risk cluster, by 30 for the final weight of virtual feature points (the larger the total weight, the higher the score), by 25 for the vehicle trajectory deviation index (the higher the deviation, the higher the score), and by 10 for (5 - minimum safe distance; if the distance is < 5 meters, take a positive value; otherwise, take 0) (the smaller the distance, the higher the score). The results of each parameter calculated according to this logic are added together to obtain the preliminary calculated value of the safety threat.

[0160] Next, the weighting coefficients are dynamically adjusted based on the real-time status:

[0161] Extract the current vehicle speed and acceleration parameters from the dynamic operating parameter dataset in step S1. If the vehicle speed exceeds the speed limit of the controlled area (e.g., the speed limit around a school is 20 km / h, and the current speed is 25 km / h), the calculation coefficient corresponding to the vehicle trajectory deviation index will be adjusted from 25 to 30 (the risk of deviation is amplified when driving at high speeds); if the absolute value of acceleration is large (e.g., rapid acceleration, acceleration 1.8), the calculation coefficient corresponding to the minimum safe distance will be adjusted from 10 to 12 (the risk of collision is higher during rapid acceleration and deceleration); if the vehicle is traveling at a constant speed and the speed is normal, all coefficients will remain unchanged; if the surrounding environment information is obtained through multi-source sensors (e.g., visual sensors, millimeter-wave radar) in the controlled area, and pedestrians are detected crossing the controlled area (e.g., students passing by the school gate), the calculation coefficient corresponding to the number of high-risk node clusters will be adjusted from 20 to 25 (the risk level increases when there are pedestrians); if there is severe weather such as rain or fog (millimeter-wave radar detects visibility of less than 50 meters), the calculation coefficients of all basic parameters will be increased by 10% overall (the perception accuracy decreases under severe weather, and risk redundancy needs to be increased); if the environment is normal (no pedestrians, clear weather), no additional coefficient adjustments will be made.

[0162] Finally, the calculated results of each parameter after dynamic adjustment are summed to obtain the corrected comprehensive score. This score is then mapped to a fixed range of 0-100 (e.g., a corrected comprehensive score of 97 corresponds to an index of 97, and a score of 35 corresponds to an index of 35), which is the vehicle safety threat index. Simultaneously, all details of the index calculation process are recorded, including the initial values ​​of the basic parameters, the calculation coefficients of each parameter (before and after adjustment), and the basis for dynamic adjustment (e.g., exceeding speed limits, pedestrian crossings), to ensure the index is traceable. If the index exceeds a preset high-risk threshold (e.g., 80), it must be marked as high-risk and requiring intervention.

[0163] In a preferred embodiment of the present invention, step S6, determining the warning level based on the vehicle safety threat index to obtain the warning level, includes:

[0164] Step S601: Based on the vehicle safety threat index, the warning level is divided into four levels: safe, attention, warning, and danger. Specifically, the vehicle safety threat index output in step S5 (usually mapped to a numerical range of 0-100, with a higher index indicating a higher collision risk) is referenced, and the safety risk tolerance of the controlled area is combined with the risk tolerance of densely populated areas such as schools and residential areas, where the risk tolerance is low and a low index is sufficient to trigger a warning. The initial threshold range for the four warning levels is then set.

[0165] Next, the core definitions and index ranges for each warning level are determined:

[0166] Safety level: The vehicle's current motion is stable with no risk of collision, corresponding to a safety threat index range of 0-30. At this level, the vehicle's trajectory conforms to a standard path, the minimum safe distance from the boundary is sufficient, and the virtual feature points are concentrated without high-risk node clusters, requiring no warning action.

[0167] Note the risk level: the vehicle exhibits a slight risk tendency (e.g., its trajectory slightly deviates from the normal path but does not approach the boundary), corresponding to a safety threat index range of 31-50. At this level, there may be a single low-weight virtual feature point (weight 0.3-0.5), or the minimum safe distance may be slightly reduced (e.g., from 2 meters to 1.5 meters), requiring the activation of a mild warning.

[0168] Warning level: The vehicle faces a moderate collision risk (e.g., its trajectory continues to deviate and gradually approaches the boundary of the controlled area), corresponding to a safety threat index range of 51-80. At this level, there are usually 1-2 high-weight virtual feature points (weight 0.8-1.0), or a small cluster of high-risk nodes (coverage < 10% of the controlled area), and the minimum safe distance is approaching the critical value (e.g., less than 1 meter), requiring the activation of a moderate intervention warning.

[0169] Danger level: The vehicle has an extremely high risk of collision (such as serious deviation from the trajectory, about to touch the boundary or has entered the danger zone), corresponding to a safety threat index range of 81-100; Under this level, the number of high-risk node clusters is ≥2, or the total weight of virtual feature points is ≥3.0, and the minimum safety distance is extremely small (such as less than 0.5 meters), requiring the activation of emergency intervention warning.

[0170] Finally, based on historical collision case data (such as minor collisions that occurred when the safety threat index reached 85 in similar control areas in the past), the boundary values ​​of the threshold ranges for each level are adjusted to ensure that the level classification can accurately match the actual risk level. At the same time, the level definition is associated with the intervention operation in the subsequent step S7 (such as the activation of physical isolation equipment corresponding to the danger level) to avoid a mismatch between the level and the intervention intensity.

[0171] Step S602: Dynamically adjust the threshold range of each warning level according to the vehicle type, characteristics of the controlled area and real-time traffic conditions. Specifically, this includes: determining the vehicle type (such as small cars, large trucks, new energy vehicles, special vehicles, etc.) from the vehicle static feature data extracted in step S1. Since the collision risk coefficients of different vehicle types vary greatly, the threshold range needs to be adjusted accordingly.

[0172] Large trucks and special vehicles (such as engineering vehicles): These vehicles have large dimensions and long braking distances, resulting in more severe collision consequences. Therefore, the overall threshold range for warning levels has been lowered. For example, the threshold for the attention level has been adjusted from 31-50 to 26-45, the warning level from 51-80 to 46-75, and the danger level from 81-100 to 76-100, ensuring that warnings are triggered at the early stages of risk for these vehicles. Small cars and new energy vehicles: These vehicles are highly agile and have short braking distances, resulting in relatively lower risk. The threshold range can remain at its initial setting or be slightly increased (e.g., the attention level is adjusted to 36-55) to avoid excessive warnings and wasted resources. Secondly, thresholds are adjusted according to the characteristics of the controlled areas. Based on the functional attributes of the controlled areas (such as school gates, residential roads, industrial park roads, highway entrances and exits), the thresholds are adjusted to adapt to the risk requirements of the areas. In densely populated areas (such as school gates and residential areas), where there is frequent human activity and the consequences of collisions are severe, the threshold range has been significantly lowered. For example, the upper limit of the safety level has been reduced from 30 to 20, and the attention level has been adjusted from 31-50 to 21-40, ensuring that even minor risks can be warned in time to protect personnel safety.

[0173] In low-density areas (such as roads within industrial parks), where there is less human activity and the primary focus is on protecting equipment and goods, the threshold range can be appropriately increased. For example, the lower limit of the warning level can be raised from 51 to 56, and the lower limit of the danger level from 81 to 86, reducing unnecessary warnings from interfering with normal traffic. Finally, the thresholds are adjusted according to real-time traffic conditions by acquiring real-time traffic data (such as traffic flow, pedestrian density, and congestion) from multi-source sensors (visual sensors, millimeter-wave radar) in the controlled area, dynamically correcting the threshold range. During congested periods / high pedestrian density, vehicle speeds are slow, but the probability of collisions is high (easily resulting in scrapes), and the risk of pedestrian intervention is significant, so the threshold range is lowered; for example, the lower limit of the attention level can be lowered from 31 to 28, allowing drivers to be promptly alerted to avoid pedestrians or other vehicles even if the vehicle's trajectory deviates slightly. During open periods / low traffic flow, where there is ample space for vehicles to travel and the risk of collisions is low, the threshold range is raised. For example, the lower limit of the warning level can be raised from 51 to 55, and the lower limit of the danger level from 81 to 85, avoiding misjudgments of high risk due to a single factor (such as instantaneous acceleration fluctuations).

[0174] Step S603: Based on historical warning data and real-time environmental factors, the warning level is verified a second time through the warning confirmation mechanism to obtain complete warning information including the warning level, risk location, and recommended handling measures. Specifically, this includes: retrieving historical warning records similar to the current scenario from the database (similar scenarios must meet the requirements of the same vehicle type, the same control area, and similar real-time traffic conditions), and extracting the correspondence between the safety threat index and the final actual risk in these historical records; if the current safety threat index is 65 (initially judged as a warning level), and in the historical data, when the index is 60-70 in the same scenario, 80% of the cases ultimately did not result in a collision, but 30% of the cases subsequently experienced an escalation of risk, then it is necessary to further verify whether the current risk has an escalation trend (such as whether the safety threat index continues to rise) to avoid directly judging it as a warning level and causing misjudgment.

[0175] If the current security threat index is 90 (initially classified as dangerous), and historical data shows that when the index is above 85 in the same scenario, 95% of cases involve risk events occurring close to the boundary, then the rationality of the dangerous level can be preliminarily confirmed, and no additional verification time is needed. Secondly, further verification is needed by combining real-time environmental factors. Real-time environmental data (such as weather conditions, lighting conditions, and temporary obstacles) is obtained through the sensor array in the controlled area to assess the impact of environmental factors on the current risk and revise the initial warning level. If the initial assessment is a warning level (index 55), but the real-time environment is rainy (millimeter-wave radar detects low visibility, increasing vehicle braking distance), then the warning level needs to be raised to a dangerous level, as rainy conditions amplify existing risks, and the original threshold range is no longer suitable. If the initial assessment is a caution level (index 38), but a temporary roadblock is added to the boundary of the controlled area in the real-time environment (visual sensors identify roadblock information), then the warning level needs to be raised to a warning level, as the roadblock reduces the vehicle's safe driving space, significantly increasing the risk.

[0176] Then, the early warning confirmation mechanism is activated to perform a secondary verification:

[0177] Compare vehicle motion data (such as position coordinates and heading angle) collected by different sensors (visual sensors, millimeter-wave radar, and lidar). If all sensor data support the current initial warning level (e.g., all show the vehicle approaching the boundary), the warning level is confirmed as valid. If a sensor's data is abnormal (e.g., lidar does not detect boundary approach while the visual sensor does), the data acquisition process is re-verified (e.g., whether there is sensor obstruction), and the level is confirmed after eliminating data errors. Continuously monitor the changing trend of the vehicle safety threat index. If the index remains within the initial level range (e.g., warning level 51-80) for 3 seconds without a downward trend, the warning level is confirmed. If the index drops rapidly within 3 seconds (e.g., from 55 to 30), it is determined to be a momentary disturbance (e.g., the driver temporarily corrects the direction), and the warning level is downgraded to a safe level to avoid false warnings. Finally, after confirming the warning level, integrate all related information to form a complete warning message containing three core components:

[0178] The warning level is determined by marking the final confirmed level (e.g., danger level) and attaching the corresponding safety threat index (e.g., 88). The risk location is marked by combining the vehicle's current location coordinates from step S3 and the spatial distribution of virtual feature points from step S4, indicating the specific risk area (e.g., near the guardrail on the east side of the controlled area, where the vehicle's current location is 0.4 meters from the guardrail), and explaining the source of the risk (e.g., severe trajectory deviation, about to touch the guardrail). Corresponding intervention operation suggestions are matched according to the warning level (e.g., for danger level, it is recommended to activate automatic bollards and send a request to the emergency management department), and these suggestions must directly correspond to the roadside equipment response plan in step S7.

[0179] In a preferred embodiment of the present invention, step S7, according to the warning level, controls the roadside equipment to perform corresponding warning and intervention operations, including:

[0180] Step S701: Activate the corresponding roadside equipment response plan according to the warning level. Specifically, the plan library stores standardized plans that correspond one-to-one with the four warning levels: safety, caution, warning, and danger. Each plan specifies the type of roadside equipment to be activated (such as information dissemination equipment, signal control equipment, physical isolation equipment, etc.), the equipment operation sequence (such as activating the audible and visual alarm first and then adjusting the traffic signal), the equipment linkage logic (such as the information dissemination equipment being synchronously associated with the vehicle location), and the backup plan for equipment failure (such as automatically switching to an adjacent screen when a certain LED screen fails).

[0181] Next, after receiving the final warning level output in step S6, the system automatically retrieves the corresponding plan from the plan library. If it is the attention level, it matches the mild information prompt plan; if it is the warning level, it matches the moderate linkage intervention plan; if it is the danger level, it matches the emergency physical isolation and emergency linkage plan; and if it is the safety level, it matches the no-intervention plan, which does not activate any roadside equipment.

[0182] Then, check the current status of each roadside device involved in the target plan (e.g., online / offline, fault / normal, power / signal strength). For example, when activating the moderate linkage intervention plan, it is necessary to confirm whether the signal control equipment is online, whether the audible and visual alarm device is fault-free, and whether the communication link is unobstructed; if a device is offline (e.g., a audible and visual alarm device loses power), the backup plan should be triggered immediately, and the same type of device in the adjacent area should be used as a substitute to ensure that the plan can be executed normally.

[0183] Finally, after confirming that the equipment is in normal condition, the system sends a start command to the target roadside equipment through a dedicated communication protocol (such as LoRa or 5G-V2X). The command includes the equipment operation parameters (such as the frequency of the audible and visual alarm and the adjustment duration of the signal timing) and the execution time node (such as completing the audible and visual alarm start within 3 seconds). The system also receives the command confirmation feedback from the equipment to ensure that each device has received and started to execute the operation, thus avoiding intervention delays caused by lost commands.

[0184] Step S702: For the attention level warning, the control information dissemination device displays warning information to remind the driver to drive safely. Specifically, this includes: based on the vehicle's current position coordinates obtained in step S3 and the spatial distribution of virtual feature points in step S4, locking the information dissemination devices in front of and to the side of the vehicle's driving path. These mainly include roadside LED warning screens (installed at high points such as light poles and monitoring poles to ensure clear visibility for the driver), voice broadcasting devices (deployed on both sides of the road median strip to adapt to different driving directions), and on-board OBU devices (if the vehicle is already connected to the system, information can be pushed synchronously), avoiding the waste of resources caused by activating devices in irrelevant areas.

[0185] Secondly, combining the vehicle trajectory deviation index from step S2 and the minimum safe distance from step S3, information content matching the actual risk of the vehicle is generated. If the vehicle trajectory deviates slightly from the normal lane, the information content is "The current trajectory deviates from the normal path, please make a slight correction to the direction." If the minimum safe distance between the vehicle and the boundary is slightly reduced, the information content is "Approaching the boundary of the controlled area, please pay attention to maintaining a safe distance." If it is nighttime or in bad weather, the information content needs to add environmental adaptation prompts such as "Poor visibility at night, it is recommended to slow down" to ensure that the information accurately reflects the risk points.

[0186] Then, the control information display equipment performs the following operations: For the LED warning screen, the system issues display instructions, determines the font color (using yellow warning color to distinguish it from other regular information), font size (ensuring clear recognition from 50 meters away), and display frequency (static display to avoid flickering and interfering with the driver's vision), and simultaneously associates the vehicle's current position with the screen's display range to ensure that the information is stably displayed when the vehicle enters the screen's visible area; For the voice broadcasting device, it issues broadcast instructions, sets the broadcast volume (10-15 decibels higher than the ambient noise level, not harsh), broadcast frequency (once every 10 seconds to avoid frequent broadcasts causing interference), and uses concise and conversational language to ensure that the driver can quickly understand.

[0187] Finally, the display status of the LED warning screen is monitored in real time by a visual sensor (such as whether the screen is lit normally and whether the content is complete), and the playback status of the voice broadcast is monitored by a sound sensor (such as whether the volume is up to standard and whether there is noise). At the same time, the millimeter-wave radar monitors whether the vehicle shows any initial response (such as slight deceleration and slight steering adjustment). If the equipment is not working properly, it immediately switches to the backup information dissemination equipment to ensure that the warning information can effectively reach the driver.

[0188] Step S703: For warning level alerts, the traffic signal control equipment is linked to adjust the traffic signal timing and activate the audible and visual alarm device. Specifically, this includes: obtaining the current signal phase (e.g., red light, green light, yellow light), remaining timing (e.g., 15 seconds left on the green light), and traffic flow distribution (e.g., 20 vehicles / minute in the east-to-west direction) of each intersection and road segment within the control area through real-time communication with the traffic signal control unit, and determining whether the current signal timing is suitable for the risk scenario (e.g., if a vehicle is near an intersection and the current green light is about to end, it may cause the vehicle to accelerate suddenly and run the light).

[0189] Next, combining the vehicle risk location from step S3 (e.g., whether it is within 50 meters of the intersection) and the trend of the safety threat index from step S5 (e.g., whether the risk continues to rise), determine the adjustment direction. If the vehicle is in the intersection's entrance lane and the current light is green, extend the green light duration by 5-10 seconds to prevent vehicles from braking suddenly or running the light when the green light ends. If the vehicle is in the intersection's exit lane and the opposite direction has a green light, shorten the opposite direction's green light duration by 3-5 seconds, while extending the green light duration in this direction to reduce the probability of oncoming vehicles meeting risky vehicles. If the controlled area is a road segment without intersections, adjust the pedestrian crossing signals on both sides of the road segment (e.g., temporarily close the pedestrian green light to prohibit pedestrians from entering the risk area).

[0190] Then, the adjustment plan is converted into instructions that the signal controller can recognize (such as extending the green light for eastbound traffic by 8 seconds and the red light for westbound traffic by 8 seconds), and sent to the signal controller through a dedicated communication link. At the same time, the controller is required to return the instruction execution result (such as the timing adjustment is complete and the new green light duration is 30 seconds). If the instruction transmission fails or the controller reports a fault, the backup plan is immediately activated, and the roadside audible and visual alarm devices are used to strengthen the reminder to surrounding vehicles to avoid the risk escalation due to the failure to adjust the signal.

[0191] Activate the audible and visual alarm device:

[0192] First, prioritize activating audible and visual alarm devices within a 30-meter radius of the vehicle's risk location, including roadside red flashing warning lights (installed on the median strips or light poles on both sides of the road) and high-volume buzzer alarms (deployed at core risk points in the controlled area, such as intersections or near boundaries). If there is a pedestrian crossing in the risk area, additional pedestrian-specific audible and visual alarm devices (such as flashing ground lights and voice prompts to be aware of and avoid vehicles) should be activated to ensure that both drivers and pedestrians can perceive the risk.

[0193] Secondly, the warning lights use a high-frequency flashing mode (3-5 flashes per second), are red (consistent with the visual perception of emergency warnings), and are set to maximum brightness (suitable for bright daylight or dark nighttime environments); the buzzer alarm uses intermittent buzzing (1 second on, 0.5 seconds off), with the volume controlled between 80-100 decibels (enough for the driver to hear clearly without causing excessive noise pollution), and the buzzing frequency is dynamically adjusted according to changes in risk (e.g., increasing the buzzing frequency when the safety threat index rises); a visual sensor confirms whether the warning lights are flashing normally, and a sound sensor confirms whether the alarm volume and frequency meet the standards; if a device malfunctions (e.g., the warning light does not illuminate), it is automatically removed from the linkage list, and a backup device is activated to ensure that there are no blind spots in the alarm coverage.

[0194] Step S704: For a hazard level warning, trigger the activation of physical isolation equipment, including automatic rising bollards and guardrails, and simultaneously send an emergency response request to the emergency management department. Specifically, based on the minimum safe distance between the vehicle and the boundary in step S3 and the risk location coordinates in step S4, locate the physical isolation equipment on the vehicle's travel path that can effectively block the risk. If the vehicle is about to crash into the boundary of the controlled area (such as a guardrail), activate the automatic rising bollard at the boundary (installed 1-2 meters inside the boundary to ensure sufficient buffer distance); if the vehicle loses control within the road segment, activate the retractable guardrail in the middle of the road segment (extending laterally to block the vehicle from further deviating from its path); if the vehicle approaches the pedestrian area, activate the smart barrier gate at the pedestrian entrance (quickly closing to prevent pedestrians from entering the risk area).

[0195] Secondly, the real-time status of the equipment is obtained through the equipment management module, including whether the hydraulic / electric system of the automatic rising bollard is normal (whether it can rise from a fully lowered state to a height of more than 1.2 meters within 3 seconds), whether the deployment mechanism of the guardrail is free from jamming (whether the deployment speed can meet the deployment requirements before the vehicle arrives), and whether the equipment power supply is sufficient (whether the battery power can support a single complete operation). If a piece of equipment malfunctions (such as insufficient hydraulic pressure on the rising bollard), the adjacent backup isolation equipment is immediately activated, and at the same time, the audible and visual alarm device is used to strengthen the reminder and make up for the temporary gap in physical isolation.

[0196] Then, the system issues a device start command and monitors the execution process: The system sends a start command to the target device, specifying the operational requirements (such as the rising column rising to a height of 1.5 meters and locking within 3 seconds and the guardrail fully unfolding within 10 seconds); at the same time, the system monitors the device's execution progress in real time through lidar and vision sensors, such as the rising height of the rising column and the unfolding extent of the guardrail, to ensure that the device completes the deployment as required; if a blockage occurs during the execution process (such as the guardrail stopping when it is unfolded to 50%), the system automatically sends a forced start command to drive the backup power unit to push the device to complete the operation.

[0197] Send an emergency response request to the emergency management department:

[0198] Extract vehicle static feature data (license plate, vehicle type, vehicle brand, license plate legality) from step S1, risk location (detailed coordinates within the control area, such as the lane 50 meters north of the east gate of XX campus) from step S3, warning level (danger level) and safety threat index from step S6, and current risk status (such as the vehicle being out of control and approaching the guardrail at a speed of 30 km / h, with the automatic bollard activated) from step S6, to ensure that the information is complete and accurate, so that emergency departments can quickly grasp the situation.

[0199] Secondly, the recipients include local traffic police departments (responsible for traffic control and vehicle interception), emergency rescue centers (responsible for possible personnel rescue and on-site handling), and management of the controlled area (such as campus security and park property management, responsible for on-site coordination); the communication method adopts multi-channel simultaneous transmission, including pushing information through a dedicated emergency communication platform, sending text messages to emergency contact persons' mobile phones, and dialing emergency duty phones (ensuring human confirmation), to avoid information loss due to single-channel failure.

[0200] Then, after sending the request, monitor the sending results of each channel in real time (such as whether the platform shows that it has been read, whether the SMS has been sent successfully, and whether the phone call has been connected); if a certain channel fails to deliver, immediately try the backup channel (if the platform fails to send, call the duty phone first); after receiving feedback from the emergency department (such as the traffic police have set off and are expected to arrive in 5 minutes), synchronize the feedback information to the system log.

[0201] Step S705 involves real-time monitoring of the effectiveness of early warning and response, and dynamic adjustment of intervention strategies based on vehicle response and changes in risk. Specifically, this includes real-time data collection via a multi-source sensor array (visual sensor, millimeter-wave radar, lidar) within the controlled area to monitor two types of core information:

[0202] Vehicle response status includes whether the vehicle decelerates (monitoring speed changes via millimeter-wave radar, such as whether it drops from 25 km / h to 15 km / h), whether it corrects its course (monitoring course angle changes via lidar, such as whether it corrects from 15° off the lane centerline to within 5°), and whether it moves away from the risk area (monitoring distance changes between the vehicle and the boundary via visual sensors, such as whether the minimum safe distance increases from 0.8 meters to 1.5 meters).

[0203] The effectiveness of roadside equipment includes whether information display equipment continues to display / broadcast normally (e.g., whether LED screens are not blacked out and whether voice is uninterrupted), whether physical isolation equipment remains in an effective state (e.g., whether bollards are stably locked at the specified height and whether guardrails are not displaced), and whether the timing of signal control equipment continues to adapt to risks (e.g., whether green light extensions still meet vehicle traffic needs).

[0204] Secondly, establish effectiveness evaluation indicators: take the change of the safety threat index as the core evaluation indicator, and set evaluation standards in combination with vehicle response and equipment execution. If the safety threat index drops from 65 at the warning level to 25 at the safety level, and the vehicle has corrected its direction and the equipment is operating normally, the handling effect is judged to be good; if the index drops from 90 at the danger level to 85, but the vehicle still does not slow down, the handling effect is judged to be insufficient; if the index does not change or continues to rise (such as from 85 to 92), the handling is judged to be ineffective.

[0205] Then, the intervention strategy is dynamically adjusted based on the assessment results:

[0206] If the handling is effective and the safety threat index remains within the safe range (0-30), gradually stop the operation of roadside equipment. First, turn off the audible and visual alarm devices, then restore the traffic signal timing to normal, and finally stop the information display equipment to avoid excessive intervention affecting normal traffic.

[0207] If the response is insufficient, strengthen existing intervention measures or activate backup plans. If vehicles at the warning level do not respond to information prompts, increase the frequency of voice broadcasts (from once every 10 seconds to once every 5 seconds) and simultaneously activate adjacent LED screens to display more prominent warning content. If physical barriers at the danger level do not completely block the risk, activate backup audible and visual alarm devices (such as installing temporary high-volume alarms) and send a supplementary request to the emergency department to expedite the response. If the response is ineffective, activate the highest level of emergency intervention. If vehicles at the danger level breach physical barriers, immediately activate all surrounding roadside equipment (such as full-area audible and visual alarms and all intersection red lights), and send a request to the emergency department for emergency interception. Simultaneously, notify other vehicles in the controlled area of ​​the risk ahead through the system platform and urge them to avoid the area immediately. Finally, record the adjustment process and results, including the reasons for the intervention strategy adjustment (such as vehicles not slowing down), the content of the adjustment (such as increasing the frequency of voice broadcasts), and the effects after the adjustment (such as vehicles starting to slow down and the index dropping to 70) in the log.

[0208] like Figure 2 As shown, an embodiment of the present invention proposes a vehicle collision detection and intervention system, comprising:

[0209] The data acquisition module is used to identify the static feature data and dynamic operating parameters of vehicles within the controlled area in real time. The static feature data includes license plate, license plate type, vehicle type, and vehicle brand; the dynamic operating parameters include vehicle speed, acceleration, heading angle, and position coordinates.

[0210] The calculation module is used to fit and predict the motion trajectory based on dynamic operating parameters, and calculate the vehicle trajectory deviation; based on the vehicle trajectory deviation and the dynamic operating parameters, it performs a collision risk assessment and calculates the minimum safe distance between the vehicle and the boundary.

[0211] The module is used to perform spatial modeling analysis on the dynamically selected vehicle trajectory curvature center point, vehicle safety envelope boundary point, and historical trajectory weighted center point based on the minimum safety distance to obtain a geometric feature model; the geometric feature model is converted into a corresponding virtual feature point set, and a dynamic evaluation structure is constructed based on the virtual feature point set;

[0212] The early warning module is used to calculate the safety threat level based on the dynamic evaluation structure to obtain the vehicle safety threat index; determine the early warning level based on the vehicle safety threat index to obtain the early warning level; and control the roadside equipment to perform corresponding early warning and intervention operations based on the early warning level.

[0213] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for vehicle collision detection and intervention, characterized in that, The method includes: The system can identify static feature data and dynamic operating parameters of vehicles within the controlled area in real time. The static feature data includes license plate, license plate type, vehicle type, and vehicle brand. The dynamic operating parameters include vehicle speed, acceleration, heading angle, and position coordinates. Based on dynamic operating parameters, motion trajectory fitting and prediction are performed, and vehicle trajectory deviation is calculated. Based on the vehicle trajectory deviation and the aforementioned dynamic operating parameters, a collision risk assessment is performed, and the minimum safe distance between the vehicle and the boundary is calculated. Based on the minimum safe distance, spatial modeling analysis is performed on the dynamically selected vehicle trajectory curvature center point, vehicle safety envelope boundary point, and historical trajectory weighted center point to obtain a geometric feature model. This includes: extracting the three-dimensional coordinate data of the vehicle trajectory curvature center point, the spatial distribution information of the vehicle safety envelope boundary point, and the spatiotemporal feature parameters of the historical trajectory weighted center point based on the minimum safe distance between the vehicle and the boundary; and normalizing the corresponding spatial coordinate data to obtain normalized coordinates of the vehicle trajectory curvature center point, normalized coordinates of the vehicle safety envelope boundary point, and normalized coordinates of the historical trajectory weighted center point; and then normalizing the coordinates of the historical trajectory weighted center point. An initial surface mesh is constructed. The coordinates of the normalized vehicle trajectory curvature center point are applied as geometric constraints to the initial surface mesh, while the coordinates of the normalized vehicle safety envelope boundary points are used as boundary conditions to obtain a spatial geometric feature surface characterizing the spatial distribution of vehicle motion. Based on this spatial geometric feature surface, the corresponding rate of curvature change and gradient direction are calculated to obtain a parametric model containing spatial geometric features. Feature extraction and dimensionality reduction are performed on the parametric model to form a geometric feature model characterizing the spatial features of vehicle motion. The geometric feature model is converted into a corresponding virtual feature point set, and a dynamic evaluation structure is constructed based on this virtual feature point set. Based on the aforementioned dynamic evaluation structure, the security threat level is calculated to obtain the vehicle security threat index; The warning level is determined based on the vehicle safety threat index. Based on the warning level, control the roadside equipment to perform corresponding warning and intervention operations.

2. The vehicle collision detection and intervention method according to claim 1, characterized in that, Real-time identification of static feature data and dynamic operating parameters of vehicles within the controlled area, wherein the static feature data includes license plate, license plate type, vehicle model and vehicle brand; The dynamic operating parameters include vehicle speed, acceleration, heading angle, and position coordinates, including: Multi-modal vehicle data is collected by a multi-source sensor array deployed at key locations in the controlled area. The multi-source sensor array includes vision sensors, millimeter-wave radar, and lidar. The vehicle images acquired by the vision sensor are preprocessed to obtain optimized vehicle image data; Vehicle features are extracted and recognized based on optimized vehicle image data, and license plate characters, license plate type, vehicle model and vehicle brand information are parsed to form a static feature dataset. Based on multi-frame point cloud data acquired simultaneously by millimeter-wave radar and lidar, point cloud registration and target clustering are performed to establish spatiotemporal trajectory information of vehicle targets. Based on the spatiotemporal trajectory information of the vehicle targets, the real-time position, motion profile and attitude information of each vehicle target are determined; By fusing the spatiotemporal trajectory information of the vehicle target with its real-time position, motion profile, and attitude information, and performing motion state estimation, the vehicle's speed, acceleration, and heading angle parameters are obtained to form a dynamic operation parameter dataset.

3. The vehicle collision detection and intervention method according to claim 2, characterized in that, Based on dynamic operating parameters, motion trajectory fitting and prediction are performed, and vehicle trajectory deviation is calculated, including: Based on the dynamic operating parameter dataset, the position coordinate data of the vehicle in continuous time series are extracted to construct the vehicle motion trajectory point set. Based on the vehicle motion trajectory point set, the order of the polynomial function used for fitting is determined; the coefficient matrix and constant term matrix of the polynomial fitting are constructed according to the time series coordinates of the trajectory point set; the least squares solution of the coefficients of each term of the polynomial is calculated by solving the regular equation formed by the coefficient matrix and the constant term matrix; the coefficients of the final fitting polynomial are determined based on the least squares solution to obtain the actual driving trajectory model that characterizes the continuous position change law of the vehicle. Based on the actual driving trajectory model of the vehicle, a spatial projection comparison is performed with the preset standard driving path to calculate the deviation distance between the two trajectories. Based on the deviation distance, a weighted normalization process is performed using dynamic operating parameters to obtain the vehicle trajectory deviation index.

4. The vehicle collision detection and intervention method according to claim 3, characterized in that, Based on the vehicle trajectory deviation and the aforementioned dynamic operating parameters, a collision risk assessment is performed, and the minimum safe distance between the vehicle and the boundary is calculated, including: Based on the vehicle trajectory deviation index and the dynamic operation parameter dataset, extract the vehicle's current position coordinates, speed, heading angle parameters, and geometric information of the control area boundary. Based on the vehicle's current position and heading angle parameters, a predictive motion spatial geometry is constructed with the vehicle position as the origin and the heading direction as the main axis. Traverse all feature line segments of the control area boundary, and calculate the perpendicular foot of the vehicle's current position point to the line containing each boundary line segment; determine whether the perpendicular foot is within the valid interval of the corresponding boundary line segment. If it is within the valid interval, record the Euclidean distance from the perpendicular foot to the vehicle's position point as a candidate distance. If the perpendicular foot is outside the valid interval, calculate the Euclidean distance from the vehicle's position point to the two endpoints of the boundary line segment and record the minimum value as a candidate distance; select the minimum spatial distance from the candidate distances corresponding to all boundary line segments; based on the vehicle's current speed and acceleration parameters, dynamically compensate and correct the minimum spatial distance to obtain dynamic distance parameters; Based on the dynamic distance parameters, a weighted fusion process is performed using the vehicle trajectory deviation index to obtain the minimum safe distance between the vehicle and the boundary.

5. The vehicle collision detection and intervention method according to claim 4, characterized in that, The geometric feature model is converted into a corresponding virtual feature point set, and a dynamic evaluation structure is constructed based on this virtual feature point set, including: Based on the geometric feature model, feature space mapping is performed to convert high-dimensional geometric features into low-dimensional feature vectors; Cluster analysis is performed on the low-dimensional feature vectors to obtain a representative set of virtual feature points, which contains key spatial location information and feature weight parameters. Based on the spatial distribution characteristics and weight parameters of the virtual feature point set, a multi-level dynamic evaluation structure is constructed, which includes a spatial relationship layer, a feature weight layer, and a risk assessment layer.

6. The vehicle collision detection and intervention method according to claim 5, characterized in that, Based on the aforementioned dynamic assessment structure, a security threat level is calculated to obtain a vehicle security threat index, including: Based on the multi-level dynamic evaluation structure, the spatial distribution features of the spatial relationship layer, the weight parameters of the feature weight layer, and the initial evaluation results of the risk assessment layer are extracted. Based on the vehicle trajectory deviation index and the minimum safe distance between the vehicle and the boundary, the various characteristic parameters are comprehensively calculated, and the weight coefficients of each characteristic parameter are dynamically adjusted according to the real-time movement status of the vehicle and changes in the surrounding environment to obtain the vehicle safety threat index.

7. A vehicle collision detection and intervention method according to claim 6, characterized in that, The warning level is determined based on the vehicle safety threat index, and includes: Based on the vehicle safety threat index, the warning levels are divided into four levels: safe, caution, warning, and danger. The threshold ranges for each warning level are dynamically adjusted based on vehicle type, characteristics of the controlled area, and real-time traffic conditions. Based on historical early warning data and real-time environmental factors, the early warning level is verified a second time through an early warning confirmation mechanism to obtain complete early warning information including the early warning level, risk location, and recommended response measures.

8. The vehicle collision detection and intervention method according to claim 7, characterized in that, Based on the warning level, control the roadside equipment to perform corresponding warning and intervention operations, including: Activate the corresponding roadside equipment response plan according to the warning level; For alert levels, the control information dissemination equipment displays warning information to remind drivers to drive safely; For warning-level alerts, the traffic signal control equipment will adjust the traffic signal timing and activate the audible and visual alarm devices. For hazard level warnings, physical isolation equipment is activated, including automatic rising bollards and guardrails, and an emergency response request is sent to the emergency management department. Real-time monitoring of the effectiveness of early warning and response, and dynamic adjustment of intervention strategies based on vehicle response and changes in risk.

9. A vehicle collision detection and intervention system, characterized in that, The system is used to perform the method as described in any one of claims 1 to 8, comprising: The data acquisition module is used to identify the static feature data and dynamic operating parameters of vehicles within the controlled area in real time. The static feature data includes license plate, license plate type, vehicle type, and vehicle brand; the dynamic operating parameters include vehicle speed, acceleration, heading angle, and position coordinates. The calculation module is used to fit and predict the motion trajectory based on dynamic operating parameters, and calculate the vehicle trajectory deviation; based on the vehicle trajectory deviation and the dynamic operating parameters, it performs a collision risk assessment and calculates the minimum safe distance between the vehicle and the boundary. The module is used to perform spatial modeling analysis on the dynamically selected vehicle trajectory curvature center point, vehicle safety envelope boundary point, and historical trajectory weighted center point based on the minimum safety distance to obtain a geometric feature model; the geometric feature model is converted into a corresponding virtual feature point set, and a dynamic evaluation structure is constructed based on the virtual feature point set; The early warning module is used to calculate the safety threat level based on the dynamic evaluation structure to obtain the vehicle safety threat index; determine the early warning level based on the vehicle safety threat index to obtain the early warning level; and control the roadside equipment to perform corresponding early warning and intervention operations based on the early warning level.

Images