Real-time traffic safety risk early warning method and system based on vehicle-road-cloud integration

Abstract

This invention discloses a real-time traffic safety risk early warning method and system based on vehicle-road-cloud integration, belonging to the field of intelligent transportation technology. The method includes: acquiring vehicle-side perception data, roadside perception data, cloud-based historical traffic data, and high-precision map data; performing spatiotemporal alignment and heterogeneous fusion of vehicle-side and roadside perception data, combining high-precision map data and cloud-based historical traffic data to generate a dynamic traffic environment digital twin, inputting it into a spatiotemporal trajectory prediction model, and outputting a future trajectory probability distribution; calculating collision risk indicators based on the future trajectory probability distribution and generating a global risk field map; generating tiered early warning information based on the global risk field map and distributing it to target vehicles to complete real-time early warning. This invention solves the technical problem that existing technologies cannot predict and quantitatively assess the future collision risks of traffic participants in advance, achieving the technical effect of forward-looking prediction and accurate early warning of traffic risks.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, specifically to a method and system for real-time early warning of traffic safety risks based on vehicle-road-cloud integration. Background Technology

[0002] In current intelligent transportation systems, traffic safety warnings largely rely on single-vehicle perception or local roadside equipment for real-time status assessment, primarily triggering alarms based on relative distance, speed, or simple threshold rules at the current moment. This approach focuses on instantaneous risk identification, lacking the ability to continuously predict the future movement trends of traffic participants and failing to comprehensively consider the interaction relationships between multiple stakeholders and overall road environment factors. Furthermore, risk assessment is typically based on point-to-point analysis, lacking a unified quantitative expression of continuous space and a global risk characterization mechanism. This results in the difficulty of timely and accurate identification and tiered warning of potential collision risks in complex traffic scenarios. Summary of the Invention

[0003] This application provides a real-time traffic safety risk early warning method and system based on vehicle-road-cloud integration, which is used to address the technical problem that existing technologies cannot predict and quantitatively assess the future collision risks of traffic participants in advance.

[0004] In view of the above problems, this application provides a method and system for real-time early warning of traffic safety risks based on vehicle-road-cloud integration.

[0005] The first aspect of this application provides a real-time traffic safety risk early warning method based on vehicle-road-cloud integration, the method comprising:

[0006] The system acquires vehicle-side perception data, roadside perception data, cloud-based historical traffic data, and high-precision map data. Based on edge computing nodes, it performs spatiotemporal alignment and heterogeneous fusion of the vehicle-side and roadside perception data, and combines this with the high-precision map data and cloud-based historical traffic data to generate a dynamic traffic environment digital twin. This dynamic traffic environment digital twin is input into a pre-trained spatiotemporal trajectory prediction model, which outputs the probability distribution of future trajectories for each traffic participant within a preset time domain. Collision risk indicators are calculated based on the future trajectory probability distribution, and the continuous space is rasterized to generate a global risk field map. Based on the global risk field map, hierarchical early warning information is generated and sent to target vehicles via the vehicle-road-cloud communication link to complete real-time early warning.

[0007] A second aspect of this application provides a real-time traffic safety risk early warning system based on vehicle-road-cloud integration, the system comprising:

[0008] The system comprises the following modules: a data acquisition module for acquiring vehicle-side perception data, roadside perception data, cloud-based historical traffic data, and high-precision map data; a digital twin generation module for spatiotemporal alignment and heterogeneous fusion of the vehicle-side perception data and roadside perception data based on edge computing nodes, and combining the high-precision map data and cloud-based historical traffic data to generate a dynamic traffic environment digital twin; a prediction module for inputting the dynamic traffic environment digital twin into a pre-trained spatiotemporal trajectory prediction model and outputting the future trajectory probability distribution of each traffic participant within a preset time domain; a rasterization processing module for calculating collision risk indicators based on the future trajectory probability distribution and rasterizing the continuous space to generate a global risk field map; and a real-time warning module for generating tiered warning information based on the global risk field map and distributing it to target vehicles via the vehicle-road-cloud communication link to achieve real-time warning.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0010] This application acquires vehicle-side perception data, roadside perception data, cloud-based historical traffic data, and high-precision map data; it performs spatiotemporal alignment and heterogeneous fusion of the vehicle-side perception data and roadside perception data based on edge computing nodes, and combines the high-precision map data and cloud-based historical traffic data to generate a dynamic traffic environment digital twin; it inputs the dynamic traffic environment digital twin into a pre-trained spatiotemporal trajectory prediction model, outputting the future trajectory probability distribution of each traffic participant within a preset time domain; it calculates collision risk indicators based on the future trajectory probability distribution, and performs rasterization processing on the continuous space to generate a global risk field map; it generates graded early warning information based on the global risk field map, and sends it to the target vehicle through the vehicle-road-cloud communication link to complete real-time early warning. This invention solves the technical problem that existing technologies cannot predict and quantify the future collision risk of traffic participants in advance. By constructing a dynamic traffic environment digital twin and combining it with a spatiotemporal trajectory prediction model to generate a future trajectory probability distribution, a global risk field is formed and graded early warnings are provided, achieving the technical effect of forward-looking prediction and accurate early warning of traffic risks. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A schematic diagram of the real-time traffic safety risk early warning method based on vehicle-road-cloud integration provided in the embodiments of this application;

[0013] Figure 2 A schematic diagram of the structure of a real-time traffic safety risk early warning system based on vehicle-road-cloud integration provided in this application embodiment.

[0014] Figure labeling: Data acquisition module 11, twin generation module 12, prediction module 13, rasterization processing module 14, real-time early warning module 15. Detailed Implementation

[0015] This application provides a real-time traffic safety risk early warning method and system based on vehicle-road-cloud integration. It addresses the technical problem that existing technologies cannot predict and quantitatively assess the future collision risks of traffic participants in advance. By constructing a dynamic traffic environment digital twin and combining it with a spatiotemporal trajectory prediction model to generate a future trajectory probability distribution, a global risk field is formed and a graded early warning is provided, thereby achieving the technical effect of forward-looking prediction and accurate early warning of traffic risks.

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0017] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0018] Example 1, as Figure 1 As shown, this application provides a real-time traffic safety risk early warning method based on vehicle-road-cloud integration, the method comprising:

[0019] Step S100: Acquire vehicle-side perception data, roadside perception data, cloud-based historical traffic data, and high-precision map data.

[0020] In this embodiment, the vehicle first continuously captures images of the road ahead using a camera mounted in front of the vehicle. The collected images are then processed in the onboard computing platform to extract the position and direction of movement information of traffic participants such as vehicles and pedestrians. Simultaneously, the vehicle reads operating status parameters such as current speed, acceleration, and steering angle from the vehicle's CAN bus, performs coordinate conversion on the recognition results, adds time stamps, and uploads the data to the edge server via the C-V2X communication link, forming structured vehicle-side perception data.

[0021] Secondly, integrated radar and vision devices are installed at key locations on the road to continuously detect traffic participants within the monitoring area and obtain information on the spatial location, speed, and category of targets in the road coordinate system. After the detection data is formatted and time-synchronized by the roadside equipment, it is transmitted back to the edge computing platform via the 5G network to form roadside perception data covering the entire road view.

[0022] Then, the traffic database of the cloud control platform retrieves road operation records within a certain time range. The records are filtered according to road segment number and time interval to extract statistical information such as historical traffic flow, average vehicle speed, congestion duration, and accident frequency. The data is then cleaned and standardized to generate a structured data table with time and space labels, thus obtaining historical traffic data in the cloud.

[0023] Finally, the vector map file of the corresponding area is downloaded from the high-precision map service platform. The information contained therein, such as the coordinates of the lane center line, lane width, road boundary, and the positions of traffic signs and traffic lights, is parsed, converted into a unified lane-level coordinate format, and stored in the database to form high-precision map data with centimeter-level positioning accuracy.

[0024] Through the above steps, vehicle-side perception data, roadside perception data, cloud-based historical traffic data, and high-precision map data are obtained respectively.

[0025] Step S200: Based on edge computing nodes, perform spatiotemporal alignment and heterogeneous fusion of the vehicle-side perception data and road-side perception data, and combine the high-precision map data and cloud-based historical traffic data to generate a dynamic traffic environment digital twin.

[0026] In this embodiment, after acquiring vehicle-side perception data and roadside perception data, edge computing nodes perform unified processing on both types of data. Specifically, based on the vehicle-side target data set uploaded via C-V2X and the roadside target data set transmitted back via 5G, the edge computing nodes use extended Kalman filtering to complete trajectory association and state estimation, forming a fused dynamic target list. The target positions, speeds, and heading angles in the dynamic target list are then mapped to the lane-level coordinate system of a high-precision map, achieving spatiotemporal alignment and heterogeneous fusion to obtain a fused and aligned dynamic target list.

[0027] Then, a multi-layer structure model is constructed by combining high-precision map data and cloud-based historical traffic data. The road infrastructure layer is constructed using lane lines, curbs, and traffic signs from the high-precision map data. The real-time traffic participant distribution layer is constructed using a fused and aligned dynamic target list. An environmental risk prior layer is constructed by extracting features of frequent congestion patterns and accident black spots from cloud-based historical traffic data. Finally, the road infrastructure layer, the real-time traffic participant distribution layer, and the environmental risk prior layer are overlaid and visualized using a lightweight 3D rendering engine to generate a dynamic digital twin of the traffic environment that can be interacted with in real time.

[0028] Furthermore, the method provided in the application embodiments, which performs spatiotemporal alignment and heterogeneous fusion of the vehicle-side perception data and the roadside perception data, further includes:

[0029] The system receives CAN bus data, millimeter-wave radar target list, and camera image recognition results uploaded by the vehicle-side via C-V2X, extracts a first target set, and obtains a vehicle-side target data set. It also receives detection data from the roadside radar-visual integrated machine transmitted via 5G, extracts a second target set, and obtains a roadside target data set. Extended Kalman filtering is used to perform trajectory association and state estimation on common targets in the first and second target sets, outputting a fused dynamic target list. The target positions, speeds, and heading angles in the dynamic target list are mapped to the lane-level coordinate system of a high-precision map to complete spatiotemporal alignment, resulting in a fused and aligned dynamic target list.

[0030] In this embodiment, after receiving CAN bus data, millimeter-wave radar target list, and camera image recognition results uploaded by the vehicle via a C-V2X communication link, the edge computing node first performs time unification processing on all data, converting the time information in each data record to the same time base and arranging them in chronological order. Then, it extracts the vehicle's own speed, acceleration, and steering angle parameters from the CAN bus data; extracts the distance, azimuth, and relative speed of each target from the millimeter-wave radar target list, and converts the distance and azimuth into planar positions in the vehicle coordinate system based on trigonometric relationships, where the lateral position is equal to the sine of the distance multiplied by the azimuth, and the longitudinal position is equal to the cosine of the distance multiplied by the azimuth; and extracts the target category and image coordinates from the camera image recognition results, and converts the image coordinates into spatial positions in the vehicle coordinate system based on camera calibration parameters. After completing the position conversion, targets from different sources are matched under the same timestamp condition. When the planar distance between two targets is less than two meters and the speed difference is less than one meter per second, they are determined to be the same traffic participant, and the corresponding data are merged to form a first target set, thereby obtaining the vehicle-side target data set.

[0031] After receiving the detection data from the roadside radar-visual integrated machine via the 5G communication link, the edge computing node also performs time unification and sorting processing, extracts the spatial position, speed, and heading angle information of each target in the road coordinate system, and deletes data with speeds exceeding twice the set road speed limit or positions exceeding the monitoring area boundary. Subsequently, the data from consecutive time moments are organized according to the target number to form a trajectory sequence, and a second target set is extracted to obtain the roadside target data set.

[0032] After obtaining the first and second target sets, trajectory association and state estimation are performed on both. First, the spatial distance between vehicle-end targets and road-end targets is calculated within the same time period. The spatial distance is equal to the square root of the sum of the squares of the lateral and longitudinal differences. When the spatial distance is less than three meters and the difference in their heading angles is less than fifteen degrees, they are identified as the same traffic participant and a corresponding relationship is established. Then, state prediction is performed. The current predicted position is calculated based on the target position and velocity at the previous moment. The predicted position is equal to the previous position plus the velocity multiplied by the time interval. Next, the difference between the predicted position and the current observed position is calculated. When the distance between the predicted and observed positions is less than two meters, the observed position is used as the update result; when the distance is greater than two meters, the predicted position is used as the correction result, thus obtaining the updated target position and velocity parameters. Through continuous prediction and correction processing, a fused dynamic target list is output.

[0033] After obtaining the fused dynamic target list, the target locations are mapped to the lane-level coordinate system of the high-precision map to complete spatiotemporal alignment. Specifically, the lane centerline coordinates are read from the high-precision map data, and the planar position of each target is projected onto the nearest lane centerline. The cumulative distance from the projection point along the lane centerline starting point to that point is calculated as the longitudinal coordinate, and the vertical distance from the target point to the lane centerline is calculated as the lateral offset. Then, adjustments are made based on the difference between the target's heading angle and the lane centerline direction to ensure that the heading angle is consistent with the lane's driving direction. Finally, the unified time information, longitudinal coordinates, lateral offset, speed, and heading angle are integrated to form the fused and aligned dynamic target list, achieving spatiotemporal alignment processing under a unified lane-level spatial reference.

[0034] Furthermore, the method provided in the application embodiments, which combines the high-precision map data with cloud-based historical traffic data to generate a dynamic traffic environment digital twin, also includes:

[0035] A static layer is constructed using lane lines, curbs, and traffic signs from a high-precision map to obtain the road infrastructure layer; a dynamic layer is constructed using a fused and aligned dynamic target list to obtain the real-time traffic participant distribution layer; a priori environmental risk layer is constructed by extracting frequently occurring congestion patterns and accident black spots from historical traffic data in the cloud; the static layer, dynamic layer, and priori environmental risk layer are overlaid and visualized using a lightweight 3D rendering engine to generate a real-time interactive dynamic traffic environment digital twin.

[0036] In this embodiment, after reading the vector files of lane lines, curbs, and traffic signs from the high-precision map data, the original latitude and longitude coordinates are first transformed into Cartesian coordinate system data. Next, the lane line points are connected sequentially to form a continuous polyline, generating the lane centerline geometry. Then, closed polygons are constructed sequentially for the curb boundary points to form the road boundary structure. Next, corresponding spatial nodes are generated for the traffic sign point data based on their spatial coordinates and attribute information. Finally, the lane centerline, road boundary, and traffic sign nodes are integrated into the same spatial data structure to construct a static layer, resulting in the road infrastructure layer, which is used to express the fixed topology and spatial geometric relationships of the road.

[0037] When constructing a dynamic layer from a merged and aligned list of dynamic targets, the system reads the lane-level longitudinal coordinates, lateral offset, velocity, and heading angle data of each target in the list. Based on the lane centerline geometry, it calculates the target's spatial position in the 3D scene, mapping the longitudinal coordinates to the cumulative distance along the lane centerline, the lateral offset to the lateral displacement relative to the centerline, and the heading angle to calculate the target's orientation. Subsequently, it loads the corresponding target entity model into the 3D scene and updates the target's position and attitude parameters at fixed time intervals to achieve continuous motion display. This constructs a dynamic layer, resulting in a real-time traffic participant distribution layer, which represents the current spatial distribution and motion state of traffic participants.

[0038] Subsequently, after accessing historical traffic data from the cloud, the historical traffic flow and average speed data are segmented and statistically analyzed according to road number. The ratio between the average speed of each lane unit and the road design speed within a set statistical period is calculated. When this ratio is continuously lower than a preset value, the corresponding area is marked as a frequently congested area. At the same time, accident records are aggregated and statistically analyzed according to spatial coordinates. The number of accidents within a set radius is counted. When the number exceeds a set threshold, the area is marked as an accident black spot feature. The frequently congested area and accident black spot features are converted into spatial grid data with risk values ​​to construct an environmental risk prior layer, which is used to express the risk distribution of the road in the historical dimension.

[0039] After the above layers are constructed, the road infrastructure layer, the real-time traffic participant distribution layer, and the environmental risk prior layer are unified under the same lane-level coordinate system for spatial overlay processing. They are aligned according to their spatial positions and written into a unified scene data structure. Then, a lightweight 3D rendering engine is called to render the overlaid scene data. The dynamic target state is updated through a real-time refresh mechanism. At the same time, the spatial distribution in the environmental risk prior layer is expressed by different color depths or transparency changes, generating a dynamic digital twin of the traffic environment that can be interacted with in real time, realizing the synchronous presentation of the real traffic environment in virtual space.

[0040] Step S300: Input the dynamic traffic environment digital twin into the pre-trained spatiotemporal trajectory prediction model and output the future trajectory probability distribution of each traffic participant within the preset time domain.

[0041] In this embodiment, after constructing the dynamic traffic environment digital twin, continuous historical trajectory data of the real-time traffic participant distribution layer is first extracted from the dynamic traffic environment digital twin. This includes the lane-level longitudinal coordinates, lateral offsets, speeds, and heading angles of each traffic participant at several consecutive time points. Simultaneously, road topology information from the road infrastructure layer is read. Then, an adjacency target interaction graph is constructed based on the spatial distance and relative position relationships between traffic participants to represent the interactive influence relationships between them. The target historical trajectory sequence and the adjacency target interaction graph are normalized and time-aligned according to a unified data format to form the model input data.

[0042] Next, the input data from the aforementioned model is fed into a pre-trained spatiotemporal trajectory prediction model. The spatiotemporal trajectory prediction model first processes the historical trajectory sequence of the target in terms of time dimension, extracting the motion trend features of each traffic participant. Then, it combines this with an adjacent target interaction graph to calculate the influence relationship between spatially neighboring targets, obtaining a fused spatiotemporal feature representation. Based on this spatiotemporal feature representation, the model recursively predicts multiple future time steps within a preset time domain, outputting several possible locations and their corresponding probability values ​​at each future time step.

[0043] Finally, the prediction results of all future time steps are integrated to obtain the probability distribution of the future trajectory of each traffic participant in the preset time domain. The sum of the probability values ​​corresponding to multiple possible positions at each time step is 1, thereby realizing the probabilistic expression of future movement trends and their uncertainties.

[0044] Furthermore, in the method provided in the application embodiments, the pre-trained spatiotemporal trajectory prediction model employs a graph attention network combined with a long short-term memory network, and its training steps include:

[0045] A sample dataset is constructed by collecting target trajectory sequences, road topology, and traffic control signal data from historical traffic scenarios. Using the target historical trajectory sequence and adjacent target interaction graph as input, and the future real trajectory as the supervision label, a trajectory probability distribution generator is trained using the negative log-likelihood loss function to obtain an initial trajectory prediction model. An adversarial training mechanism is introduced, and a discriminator is used to distinguish between the generated trajectory and the real trajectory for adversarial optimization. The process is iterated until the Nash equilibrium state is reached to obtain a converged spatiotemporal trajectory prediction model.

[0046] In this embodiment, target trajectory sequences, road topology, and traffic control signal data from historical traffic scenarios are first collected. The target trajectory sequence is generated by resampling vehicle trajectory points at fixed time intervals to form time-series data containing continuous position coordinates, speed, and heading angle. The road topology is generated by parsing the lane connection relationships in a high-precision map to create a lane adjacency matrix. The traffic control signal data is generated by reading the traffic light timing table to create a time-labeled sequence. Subsequently, the three types of data are aligned according to a unified time axis, and fixed-length historical trajectory segments are extracted using a sliding time window as input samples. The real trajectories after the corresponding time period are used as supervision labels to construct a sample dataset.

[0047] During model training, the target historical trajectory sequence and the adjacent target interaction graph are used as inputs. The adjacent target interaction graph generates an adjacency matrix by calculating the spatial distance between traffic participants and setting an adjacency threshold, which is used to represent the interaction relationship between traffic participants. The target historical trajectory sequence is input into a long short-term memory network, and motion features in the time dimension are extracted through time recursion. The adjacent target interaction graph is input into a graph attention network, and spatial interaction features are calculated by assigning attention weights to adjacent nodes. Subsequently, the temporal and spatial features are concatenated and fused, and input into a trajectory probability distribution generator to output position prediction parameters for multiple future time steps. In the optimization process, a negative log-likelihood loss function is used to perform probability matching calculations between the trajectory probability distribution output by the model and the future real trajectory. The model parameters are updated using the gradient descent method to obtain the initial trajectory prediction model.

[0048] After obtaining the initial trajectory prediction model, an adversarial training mechanism is introduced to improve the realism of the prediction results. Specifically, a discriminator network is constructed, and the generated predicted trajectory and the real trajectory are respectively input into the discriminator. The discriminator outputs a discrimination result indicating its degree of realism. During the training process, two optimization steps are performed alternately: one step fixes the discriminator parameters and updates the trajectory probability distribution generator to make the generated trajectory closer to the real trajectory distribution; the other step fixes the trajectory probability distribution generator parameters and updates the discriminator to improve its ability to distinguish between the generated trajectory and the real trajectory. By iteratively updating the parameters of both, the generated trajectory gradually approximates the real trajectory in terms of statistical distribution until the trajectory probability distribution generator and the discriminator reach a Nash equilibrium, ultimately obtaining a converged spatiotemporal trajectory prediction model.

[0049] Step S400: Calculate the collision risk index based on the future trajectory probability distribution, and perform rasterization processing on the continuous space to generate a global risk field map.

[0050] In this embodiment of the application, when calculating the collision risk index based on the probability distribution of future trajectories, the possible positions of any two vehicles in a preset time domain are first jointly analyzed based on the probability distribution of future trajectories. By matching the future positions of the two vehicles in time, it is determined whether they enter the same spatial region, and the time difference between the arrival of the two vehicles in the spatial region is calculated. The minimum time difference is selected as the time risk parameter. At the same time, combined with the current relative speed of the two vehicles and the vehicle deceleration limit, the minimum deceleration required to avoid contact under the time risk parameter is calculated to obtain the dynamic risk parameter. Subsequently, the time risk parameter and the dynamic risk parameter are dimensionless and fused to generate a collision risk index for quantifying the potential collision degree between the two vehicles.

[0051] After obtaining the collision risk index, the continuous space covered by the digital twin of the dynamic traffic environment is rasterized, dividing the continuous space into several raster units to construct a risk calculation grid. For each raster unit, the probability of it being occupied by any traffic participant within a preset time domain is calculated based on the probability distribution of future trajectories, and adjusted in conjunction with factors such as traffic participant type, speed, and size to obtain an individual risk contribution value. Subsequently, the individual risk contribution values ​​of each traffic participant within the same raster unit are superimposed to calculate the joint risk value of that raster unit, which is expressed in a spatial distribution form, and finally a global risk field map reflecting the overall regional risk distribution is generated.

[0052] Furthermore, in the method provided in the application embodiments, calculating the collision risk index based on the future trajectory probability distribution further includes:

[0053] Based on the probability distribution of the future trajectories of the two vehicles, multiple possible position-time pairs are sampled, the time difference of arrival in the same spatial region is calculated, and the minimum time difference is obtained as the collision time to obtain the time risk parameter; based on the current relative speed and deceleration limit of the two vehicles, the minimum deceleration required to avoid the collision is calculated to obtain the collision avoidance deceleration requirement to obtain the dynamic risk parameter; the time risk parameter and the dynamic risk parameter are dimensionless and then weighted and fused to generate the collision risk index.

[0054] In this embodiment, when sampling multiple possible location-time points based on the probability distribution of the future trajectories of the two vehicles, a fixed time step is first selected within a preset time domain, and the preset time domain is discretized into multiple future moments. Then, for the probability distribution of the future trajectories of vehicle one and vehicle two at each future moment, multiple candidate location points are generated using a probability sampling method. That is, each candidate location in the probability distribution of the future trajectory is arranged according to its probability value, the cumulative sum of the probability values ​​is calculated, and a random number between zero and one is generated and falls into the cumulative probability interval to determine a sampling location. Multiple candidate location points are obtained by repeating this sampling process. Finally, the candidate locations of vehicle one and vehicle two obtained from each sampling are combined with the corresponding future moments to form multiple location-time point pairs.

[0055] When calculating the time difference of arrival in the same spatial region and obtaining the minimum time difference as the collision time, the spatial distance between the two vehicles is first calculated for each location time point. The spatial distance is obtained by adding the squares of the lateral and longitudinal differences between the candidate positions of vehicle one and vehicle two in the planar coordinate system and taking the square root. Then, the spatial distance is compared with a preset spatial threshold, which is the minimum safe distance corresponding to the sum of half the width of the two vehicles and half the length of the two vehicles. When the spatial distance is less than the preset spatial threshold, it is determined that the two vehicles have entered the same spatial region. For all cases that meet the condition of entering the same spatial region, the future time corresponding to vehicle one and vehicle two is recorded, and the time difference is calculated as the absolute difference between the two future times. By traversing all recorded time differences and selecting the minimum value, the collision time is determined, and the collision time is converted into a time risk parameter. The time risk parameter is the reciprocal of the collision time, so that the smaller the collision time, the larger the time risk parameter.

[0056] Then, based on the current relative speed and deceleration limit of the two vehicles, the minimum deceleration required to avoid a collision is calculated, and the collision avoidance deceleration requirement is obtained. First, the relative speed is calculated from the current speed of the two vehicles. The relative speed is the difference between the speed of vehicle two and the speed of vehicle one, and the absolute value is taken. Then, the vehicle deceleration limit is read as the maximum achievable braking capacity. Next, the collision time is used as the available reaction time, and the minimum deceleration is calculated according to the uniform deceleration relationship. The minimum deceleration is equal to the relative speed divided by the collision time. The minimum deceleration is compared with the deceleration limit to obtain the collision avoidance deceleration requirement. When the minimum deceleration is less than or equal to the deceleration limit, the collision avoidance deceleration requirement is taken as the minimum deceleration. When the minimum deceleration is greater than the deceleration limit, the collision avoidance deceleration requirement is taken as the deceleration limit. This is used to represent the braking demand intensity of the vehicle within its power capacity range, and the collision avoidance deceleration requirement is used as a power risk parameter.

[0057] Finally, when generating the collision risk index by weighted fusion after dimensionless processing of the time risk parameters and dynamic risk parameters, the time risk parameters are first dimensionlessized by selecting a preset time threshold as a benchmark and dividing the time risk parameters by the reciprocal of the preset time threshold to obtain the normalized time risk value. Then, the dynamic risk parameters are dimensionlessized by dividing them by the deceleration limit to obtain the normalized dynamic risk value. Next, technical experts pre-set time weighting coefficients and dynamic weighting coefficients, with the sum of the two being one. Finally, the collision risk index is calculated, which equals the normalized time risk value multiplied by the time weighting coefficient plus the normalized dynamic risk value multiplied by the dynamic weighting coefficient. This yields a collision risk index that simultaneously reflects the urgency of time and the degree of braking demand.

[0058] Furthermore, in the method provided in the application embodiments, the process of rasterizing a continuous space to generate a global risk field map further includes:

[0059] The continuous space covered by the digital twin of the dynamic traffic environment is discretized into grid cells to obtain a risk calculation grid. The probability of each grid cell being occupied by the future trajectory of any traffic participant is calculated, and the risk contribution value is obtained by combining the traffic participant type, speed and size with potential energy weighting. The individual risk contribution values ​​of each traffic participant are superimposed to calculate the joint risk value of each grid cell and visualized in the form of equipotential lines to generate a global risk field map.

[0060] In this embodiment, when discretizing the continuous space covered by the dynamic traffic environment digital twin into grid cells to obtain the risk calculation grid, the spatial coverage area determined by the road infrastructure layer is first read from the dynamic traffic environment digital twin, and the minimum longitudinal coordinate, maximum longitudinal coordinate, minimum lateral coordinate, and maximum lateral coordinate of this area are obtained. Then, the grid side length is set as the grid resolution, and the grid side length is accumulated segment by segment in the longitudinal direction from the minimum longitudinal coordinate until the maximum longitudinal coordinate to obtain the longitudinal dividing line sequence. In the lateral direction, the grid side length is accumulated segment by segment from the minimum lateral coordinate until the maximum lateral coordinate to obtain the lateral dividing line sequence. Then, the rectangular area enclosed by two adjacent longitudinal dividing lines and two adjacent lateral dividing lines is defined as a grid cell, and each grid cell is assigned a unique number and boundary coordinates. The set of all grid cells forms the risk calculation grid.

[0061] Next, when calculating the probability of each grid cell being occupied by any traffic participant's future trajectory, the probability distribution of the traffic participant's future trajectory is first read, and the preset time domain is discretized into multiple future time steps with a time interval of 0.5 seconds. Multiple possible positions and corresponding probability values ​​are obtained at each future time step. Then, a grid-based judgment is performed on each possible position, that is, it is determined whether its vertical coordinate is located between the vertical boundaries of a grid cell, and whether its horizontal coordinate is located between the horizontal boundaries of the grid cell. If the conditions are met, the probability value corresponding to the possible position is added to the occupancy probability of the grid cell at that time step. The above steps are repeated for all future time steps, and the occupancy probabilities of each time step are added together to obtain the overall probability that the grid cell is occupied by the traffic participant's future trajectory.

[0062] Before performing calculations based on traffic participant type, speed, and size, the relevant parameters are normalized to ensure that different dimensions can be directly multiplied. First, set a base type value, for example, 1.5 for heavy trucks, 1.0 for small passenger cars, and 2.0 for pedestrians. Then, divide the base type value by the set maximum base type value of 2.0 to obtain the normalized type coefficient, which ranges from 0 to 1. Next, read the current speed and calculate the speed coefficient, which is equal to the current speed divided by the road speed limit. For example, if the speed limit is 20 meters per second and the current speed is 10 meters per second, the speed coefficient is 0.5. If the speed limit is exceeded, it is 1.0. Then, read the size parameter, which is equal to the sum of the length and width. For example, if the length is 4 meters and the width is 1.8 meters, the size parameter is 5.8 meters. Set the maximum reference size to 8 meters, then the size coefficient is 5.8 divided by 8, which is approximately 0.725. The above type coefficient, speed coefficient, and size coefficient are all between 0 and 1. Then, multiply the occupancy probability by the normalized type coefficient, normalized speed coefficient, and normalized size coefficient to obtain the individual risk contribution value of the traffic participant in that grid cell.

[0063] Then, the individual risk contribution values ​​of each traffic participant are superimposed, and the joint risk value of each grid cell is calculated. Specifically, for each grid cell in the risk calculation grid, the individual risk contribution values ​​of all traffic participants in that grid cell are added one by one to obtain the joint risk value of that grid cell; if the joint risk value is greater than 1, it is normalized, and the joint risk value is equal to the current value divided by the maximum joint risk value in that area, thus ensuring that the joint risk value is between 0 and 1; the above calculation is repeated for all grid cells to form a joint risk value matrix covering the entire area.

[0064] Finally, when visualizing the global risk field map using equipotential lines, the following steps are taken: First, set the equipotential levels, such as 0.2, 0.4, 0.6, and 0.8, as the joint risk value thresholds. Then, find the locations in the risk calculation grid where the joint risk value between adjacent grid cells crosses a certain threshold. Calculate the coordinates of the crossing points using linear interpolation and connect the crossing points in spatial order to form a continuous curve, obtaining the equipotential lines corresponding to the thresholds. Finally, overlay the equipotential lines onto the dynamic traffic environment digital twin to generate the global risk field map.

[0065] Step S500: Generate hierarchical early warning information based on the global risk field map matching, and send it to the target vehicle through the vehicle-road-cloud communication link to complete the real-time early warning.

[0066] In this embodiment, the location of the grid cell corresponding to the target vehicle is first determined in the lane-level coordinate system. The joint risk value of the grid cell is read and used as the risk field value for graded early warning determination. According to the preset risk threshold range, the risk field value is compared with each threshold to determine the corresponding early warning level and generate graded early warning information containing risk level identifier, risk location coordinates, time information, and handling suggestions. When the risk field value reaches a higher threshold, a speed control strategy or path guidance strategy is generated simultaneously and processed by the edge computing node. After consistency verification by the cloud control platform, the strategy is sent to the target vehicle through the C-V2X communication channel in the vehicle-road-cloud communication link. The vehicle receives and parses the early warning information and executes corresponding prompts or control operations, thereby completing the process of graded early warning generation, information distribution, and real-time early warning response based on the global risk field map.

[0067] Furthermore, in the method provided in the application embodiments, generating hierarchical early warning information based on the global risk field map matching further includes:

[0068] Multiple warning levels are preset, each corresponding to different risk threshold ranges, levels of detail in the warning content, and warning push ranges. The rules for determining the graded warnings are established. When the risk field value of the area where the target vehicle is located exceeds the first threshold, a level 1 warning is generated. When the risk field value exceeds the second threshold, a level 2 warning is generated. When the risk field value exceeds the third threshold, a level 3 warning is generated, and speed control or path guidance strategies are sent to edge computing nodes. After verification by the cloud control platform, the warnings are pushed to the vehicle to complete the high-level risk warning.

[0069] In this embodiment, multiple warning levels are first preset, each corresponding to a different risk threshold range, level of detail in the warning content, and warning push range, to determine the tiered warning judgment rules. Specifically, a warning level field is established in the parameter table, and three warning level identifiers are written: Level 1 warning information, Level 2 warning information, and Level 3 warning information. Simultaneously, three risk threshold parameters are written: a first threshold, a second threshold, and a third threshold, with a constraint that the first threshold is less than the second threshold and the second threshold is less than the third threshold. Then, corresponding risk threshold range judgment conditions are written for each warning level. For example, Level 1 warning information corresponds to a risk field value greater than the first threshold but not exceeding the second threshold; Level 2 warning information corresponds to a risk field value greater than the second threshold but not exceeding the third threshold; and Level 3 warning information corresponds to a risk field value greater than the third threshold. Here, the risk field value is the ratio of the target vehicle's grid cell in the global risk field map to the risk field value. The system first assigns a joint risk value; then, it writes a field for the level of detail of the warning content for each warning level. For example, a level 1 warning includes a risk warning and risk direction information; a level 2 warning includes a risk warning, risk location coordinates, and suggested deceleration information; and a level 3 warning includes a risk warning, risk location coordinates, suggested deceleration information, and speed control strategy or path guidance strategy fields. Finally, it writes a field for the warning push range for each warning level. For example, a level 1 warning push range is limited to the target vehicle; a level 2 warning push range is limited to the target vehicle and vehicles within a certain distance behind it; and a level 3 warning push range is limited to the target vehicle and related vehicles within the same risk area. This forms a directly callable hierarchical warning judgment rule.

[0070] When the risk field value of the area where the target vehicle is located exceeds the first threshold, the grid cell number where the target vehicle is located is located in the lane-level coordinate system, the risk field value of the grid cell is read, and the risk field value is compared with the first threshold. If the risk field value is greater than the first threshold, the first-level warning information generation process is triggered. Then, according to the warning content detail field, the content template required for the first-level warning information is read from the parameter table, the target vehicle identifier, risk field value, risk area location coordinates and timestamp are written into the warning message field, the first-level warning information is generated, and the recipient of the first-level warning information is determined as the target vehicle according to the warning push range field.

[0071] When the risk field value exceeds the second threshold, the risk field value is read under the same grid cell number, and the risk field value is compared with the second threshold. If the risk field value is greater than the second threshold, the secondary warning information generation process is triggered. Then, the content template corresponding to the secondary warning information is read from the parameter table, and the target vehicle identification, risk field value, risk area location coordinates, risk direction information and suggested deceleration information are written into the warning message field to generate the secondary warning information. Then, the receiving target range of the secondary warning information is determined according to the warning push range field, and the vehicle identification of the target vehicle and vehicles within a certain distance behind it are written into the receiving list to complete the determination of the push target of the secondary warning information.

[0072] When the risk field value exceeds the third threshold, the risk field value of the grid cell containing the target vehicle is read and compared with the third threshold. If the risk field value is greater than the third threshold, the three-level warning information generation process is triggered. Subsequently, the content template corresponding to the three-level warning information is read from the parameter table, and the target vehicle identifier, risk field value, risk area location coordinates, risk direction information, and suggested deceleration information are written into the warning message field to generate the three-level warning information. At the same time, a speed control strategy or a path guidance strategy is generated. The speed control strategy calculates the target deceleration range by measuring the difference between the target vehicle's current speed and the target safe speed, and provides the target speed value. The path guidance strategy reads the adjacent lane connections in the high-precision map data. The system determines the available lane change direction and provides the target lane number or target travel route. The speed control strategy or route guidance strategy is encapsulated into a strategy message and sent to the edge computing node. The edge computing node performs field integrity verification and vehicle identifier matching verification on the strategy message before uploading it to the cloud control platform. The cloud control platform performs strategy verification, including threshold consistency checks and conflict checks. Once it confirms that the strategy does not conflict with the current traffic control signal data and does not exceed the vehicle's deceleration limit, an executable strategy is formed. Finally, the cloud control platform pushes the level-three warning information and the executable strategy to the vehicle via the vehicle-road-cloud communication link. The vehicle parses the information and executes speed control or route guidance, thus completing the high-level risk warning.

[0073] In summary, the embodiments of this application have at least the following technical effects:

[0074] This application acquires vehicle-side perception data, roadside perception data, cloud-based historical traffic data, and high-precision map data; it performs spatiotemporal alignment and heterogeneous fusion of the vehicle-side perception data and roadside perception data based on edge computing nodes, and combines the high-precision map data and cloud-based historical traffic data to generate a dynamic traffic environment digital twin; it inputs the dynamic traffic environment digital twin into a pre-trained spatiotemporal trajectory prediction model, outputting the future trajectory probability distribution of each traffic participant within a preset time domain; it calculates collision risk indicators based on the future trajectory probability distribution, and performs rasterization processing on the continuous space to generate a global risk field map; it generates graded early warning information based on the global risk field map, and sends it to the target vehicle through the vehicle-road-cloud communication link to complete real-time early warning. This invention solves the technical problem that existing technologies cannot predict and quantify the future collision risk of traffic participants in advance. By constructing a dynamic traffic environment digital twin and combining it with a spatiotemporal trajectory prediction model to generate a future trajectory probability distribution, a global risk field is formed and graded early warnings are provided, achieving the technical effect of forward-looking prediction and accurate early warning of traffic risks.

[0075] Example 2 is based on the same inventive concept as the vehicle-road-cloud integrated real-time traffic safety risk early warning method in the previous examples, such as... Figure 2 As shown, this application provides a real-time traffic safety risk early warning system based on vehicle-road-cloud integration. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0076] The data acquisition module 11 is used to acquire vehicle-side perception data, roadside perception data, cloud-based historical traffic data, and high-precision map data; the digital twin generation module 12 is used to perform spatiotemporal alignment and heterogeneous fusion of the vehicle-side perception data and roadside perception data based on edge computing nodes, and combine the high-precision map data and cloud-based historical traffic data to generate a dynamic traffic environment digital twin; the prediction module 13 is used to input the dynamic traffic environment digital twin into a pre-trained spatiotemporal trajectory prediction model and output the future trajectory probability distribution of each traffic participant within a preset time domain; the rasterization processing module 14 is used to calculate collision risk indicators based on the future trajectory probability distribution and perform rasterization processing on the continuous space to generate a global risk field map; the real-time warning module 15 is used to generate graded warning information based on the global risk field map and send it to the target vehicle through the vehicle-road-cloud communication link to complete the real-time warning.

[0077] Furthermore, the system is also used to implement the following functions:

[0078] The system receives CAN bus data, millimeter-wave radar target list, and camera image recognition results uploaded by the vehicle-side via C-V2X, extracts a first target set, and obtains a vehicle-side target data set. It also receives detection data from the roadside radar-visual integrated machine transmitted via 5G, extracts a second target set, and obtains a roadside target data set. Extended Kalman filtering is used to perform trajectory association and state estimation on common targets in the first and second target sets, outputting a fused dynamic target list. The target positions, speeds, and heading angles in the dynamic target list are mapped to the lane-level coordinate system of a high-precision map to complete spatiotemporal alignment, resulting in a fused and aligned dynamic target list.

[0079] Furthermore, the system is also used to implement the following functions:

[0080] A static layer is constructed using lane lines, curbs, and traffic signs from a high-precision map to obtain the road infrastructure layer; a dynamic layer is constructed using a fused and aligned dynamic target list to obtain the real-time traffic participant distribution layer; a priori environmental risk layer is constructed by extracting frequently occurring congestion patterns and accident black spots from historical traffic data in the cloud; the static layer, dynamic layer, and priori environmental risk layer are overlaid and visualized using a lightweight 3D rendering engine to generate a real-time interactive dynamic traffic environment digital twin.

[0081] Furthermore, the system is also used to implement the following functions:

[0082] A sample dataset is constructed by collecting target trajectory sequences, road topology, and traffic control signal data from historical traffic scenarios. Using the target historical trajectory sequence and adjacent target interaction graph as input, and the future real trajectory as the supervision label, a trajectory probability distribution generator is trained using the negative log-likelihood loss function to obtain an initial trajectory prediction model. An adversarial training mechanism is introduced, and a discriminator is used to distinguish between the generated trajectory and the real trajectory for adversarial optimization. The process is iterated until the Nash equilibrium state is reached to obtain a converged spatiotemporal trajectory prediction model.

[0083] Furthermore, the system is also used to implement the following functions:

[0084] Based on the probability distribution of the future trajectories of the two vehicles, multiple possible position-time pairs are sampled, the time difference of arrival in the same spatial region is calculated, and the minimum time difference is obtained as the collision time to obtain the time risk parameter; based on the current relative speed and deceleration limit of the two vehicles, the minimum deceleration required to avoid the collision is calculated to obtain the collision avoidance deceleration requirement to obtain the dynamic risk parameter; the time risk parameter and the dynamic risk parameter are dimensionless and then weighted and fused to generate the collision risk index.

[0085] Furthermore, the system is also used to implement the following functions:

[0086] The continuous space covered by the digital twin of the dynamic traffic environment is discretized into grid cells to obtain a risk calculation grid. The probability of each grid cell being occupied by the future trajectory of any traffic participant is calculated, and the risk contribution value is obtained by combining the traffic participant type, speed and size with potential energy weighting. The individual risk contribution values ​​of each traffic participant are superimposed to calculate the joint risk value of each grid cell and visualized in the form of equipotential lines to generate a global risk field map.

[0087] Furthermore, the system is also used to implement the following functions:

[0088] Multiple warning levels are preset, each corresponding to different risk threshold ranges, levels of detail in the warning content, and warning push ranges. The rules for determining the graded warnings are established. When the risk field value of the area where the target vehicle is located exceeds the first threshold, a level 1 warning is generated. When the risk field value exceeds the second threshold, a level 2 warning is generated. When the risk field value exceeds the third threshold, a level 3 warning is generated, and speed control or path guidance strategies are sent to edge computing nodes. After verification by the cloud control platform, the warnings are pushed to the vehicle to complete the high-level risk warning.

[0089] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0090] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A real-time traffic safety risk early warning method based on vehicle-road-cloud integration, characterized in that, The method includes: Acquire vehicle-side perception data, roadside perception data, cloud-based historical traffic data, and high-precision map data; Based on edge computing nodes, the vehicle-side perception data and roadside perception data are spatiotemporally aligned and heterogeneously fused. Combined with the high-precision map data and cloud-based historical traffic data, a dynamic traffic environment digital twin is generated. The dynamic traffic environment digital twin is input into a pre-trained spatiotemporal trajectory prediction model, which outputs the probability distribution of future trajectories of each traffic participant within a preset time domain. The collision risk index is calculated based on the probability distribution of the future trajectory, and the continuous space is rasterized to generate a global risk field map. Based on the global risk field map matching, hierarchical early warning information is generated and sent to the target vehicle through the vehicle-road-cloud communication link to complete real-time early warning.

2. The real-time traffic safety risk early warning method based on vehicle-road-cloud integration according to claim 1, characterized in that, The spatiotemporal alignment and heterogeneous fusion of the vehicle-side perception data and the roadside perception data include: The system receives CAN bus data, millimeter-wave radar target list, and camera image recognition results uploaded by the vehicle via C-V2X, extracts the first target set, and obtains the vehicle-side target data set. Receive the detection data from the radar-visual integrated machine transmitted back via 5G at the roadside, extract the second target set, and obtain the roadside target data set; Extended Kalman filtering is used to perform track association and state estimation on common targets in the first and second target sets, and a fused dynamic target list is output. The target positions, speeds, and heading angles in the dynamic target list are mapped to the lane-level coordinate system of the high-precision map to complete spatiotemporal alignment, resulting in a fused and aligned dynamic target list.

3. The real-time traffic safety risk early warning method based on vehicle-road-cloud integration according to claim 1, characterized in that, By combining the high-precision map data with historical traffic data from the cloud, a dynamic digital twin of the traffic environment is generated, including: A static layer is constructed using lane lines, curbs, and traffic signs from a high-precision map to obtain the road infrastructure layer; A dynamic layer is constructed using the merged and aligned dynamic target list to obtain the real-time traffic participant distribution layer; Based on historical traffic data from the cloud, frequently occurring congestion patterns and accident hotspots are extracted to construct an environmental risk prior layer. The static layer, dynamic layer, and environmental risk prior layer are overlaid and visualized using a lightweight 3D rendering engine to generate a dynamic digital twin of the traffic environment that can be interacted with in real time.

4. The real-time traffic safety risk early warning method based on vehicle-road-cloud integration according to claim 1, characterized in that, The pre-trained spatiotemporal trajectory prediction model employs a graph attention network combined with a long short-term memory network. The training steps include: Collect target trajectory sequences, road topology, and traffic control signal data from historical traffic scenarios to construct a sample dataset; Using the target's historical trajectory sequence and the interaction graph of adjacent targets as input, and the future real trajectory as the supervision label, a trajectory probability distribution generator is trained using the negative log-likelihood loss function to obtain the initial trajectory prediction model. An adversarial training mechanism is introduced, and a discriminator is used to distinguish between generated trajectories and real trajectories for adversarial optimization. The process is iterated until the Nash equilibrium state is reached, resulting in a convergent spatiotemporal trajectory prediction model.

5. The real-time traffic safety risk early warning method based on vehicle-road-cloud integration according to claim 1, characterized in that, The collision risk index is calculated based on the future trajectory probability distribution, including: Based on the probability distribution of the future trajectories of the two vehicles, multiple possible position-time pairs are sampled, the time difference of arrival in the same spatial region is calculated, and the minimum time difference is obtained as the collision time to obtain the time risk parameter. Calculate the minimum deceleration required to avoid a collision based on the current relative speed and deceleration limit of the two vehicles, obtain the collision avoidance deceleration requirement, and obtain the dynamic risk parameters; The time risk parameters and dynamic risk parameters are dimensionless and then weighted and fused to generate a collision risk index.

6. The real-time traffic safety risk early warning method based on vehicle-road-cloud integration according to claim 1, characterized in that, The continuous space is rasterized to generate a global risk field map, including: The continuous space covered by the digital twin of the dynamic traffic environment is discretized into grid cells to obtain the risk calculation grid; For each grid cell, the probability that it will be occupied by the future trajectory of any traffic participant is calculated, and the potential energy weighting is combined with the traffic participant type, speed and size to obtain the individual risk contribution value. By superimposing the individual risk contribution values ​​of each traffic participant, the joint risk value of each grid cell is calculated and visualized in the form of equipotential lines to generate a global risk field map.

7. The real-time traffic safety risk early warning method based on vehicle-road-cloud integration according to claim 1, characterized in that, Based on the global risk field map matching, hierarchical early warning information is generated, including: Multiple warning levels are preset, each corresponding to different risk threshold ranges, levels of detail in the warning content, and warning push scope, and the rules for determining the graded warnings are established. When the risk field value of the area where the target vehicle is located exceeds the first threshold, a Level 1 warning message is generated; When the risk field value exceeds the second threshold, a level-two early warning message is generated; When the risk field value exceeds the third threshold, a level 3 early warning message is generated, and speed control or path guidance strategies are sent to the edge computing node. After verification by the cloud control platform, the message is pushed to the vehicle to complete the high-level risk warning.

8. A real-time traffic safety risk early warning system based on vehicle-road-cloud integration, characterized in that: The system is used to execute the real-time traffic safety risk early warning method based on vehicle-road-cloud integration as described in any one of claims 1-7, and the system includes: The data acquisition module is used to acquire vehicle-side perception data, roadside perception data, cloud-based historical traffic data, and high-precision map data. The twin generation module is used to perform spatiotemporal alignment and heterogeneous fusion of the vehicle-side perception data and road-side perception data based on edge computing nodes, and to generate a dynamic traffic environment digital twin by combining the high-precision map data and cloud-based historical traffic data. The prediction module is used to input the digital twin of the dynamic traffic environment into a pre-trained spatiotemporal trajectory prediction model and output the probability distribution of the future trajectories of each traffic participant within a preset time domain. The rasterization module is used to calculate the collision risk index based on the probability distribution of the future trajectory, and to perform rasterization processing on the continuous space to generate a global risk field map. The real-time early warning module is used to generate hierarchical early warning information based on the global risk field map matching, and send it to the target vehicle through the vehicle-road-cloud communication link to complete the real-time early warning.

Images