Expressway multi-source vehicle collision risk prediction and warning method

Abstract

The present application relates to the field of alarm signal device, specifically to a highway multi-source vehicle collision risk prediction alarm method, receiving multi-vehicle trajectory time series data fused and generated by roadside multi-source sensing equipment; constructing a dynamic space-time topology with vehicles as nodes and vehicle distance and relative speed as edge weights; extracting space-time interaction features of vehicle groups by using space-time graph convolution network, and inputting the features into a non-cooperative game model to calculate the game equilibrium trajectory of each vehicle in the strategy space; calculating the space-time conflict degree of the multi-vehicle game trajectory in the prediction time domain, generating collision risk entropy based on the time integral of the space-time conflict degree; when the collision risk entropy exceeds the dynamic adaptive threshold, triggering the generation of a hierarchical alarm instruction carrying intervention priority and spatial orientation information and outputting. The present application overcomes the deviation caused by single vehicle prediction ignoring group game effect, realizes the quantification of multi-vehicle collision risk, solves the prediction lag and alarm failure problem in the strong interaction scene, and improves the system reliability.

Description

Technical Field

[0001] This invention relates to the field of alarm signal devices, specifically a method for predicting and alarming multi-source vehicle collision risks on highways. Background Technology

[0002] Highway vehicle collision risk prediction and warning systems typically rely on roadside sensing devices to acquire traffic flow data and use this data to predict vehicle trajectories and assess risks. In conventional implementations, after receiving vehicle position and speed data from radar or visual sensors, the system often employs extrapolation prediction methods based on single-vehicle kinematic models. These methods treat the target vehicle as an isolated entity, using its historical position coordinates and velocity vectors to establish motion equations based on constant speed or constant acceleration assumptions, thus extrapolating the vehicle's trajectory over a short period. At the multi-source data fusion level, existing technologies often employ simple coordinate transformations and data-level stitching, failing to perform deep alignment processing for temporal deviations in the perceived data, resulting in inherent jitter in the trajectory time-series data input to the prediction model.

[0003] In the collision risk assessment and alarm process, existing technologies primarily rely on threshold-based mechanisms based on physical geometric relationships. When predicted vehicle trajectories overlap on a two-dimensional plane, or when the calculated predicted collision time is lower than a preset fixed threshold, the system determines a collision risk and triggers an alarm. The alarm signal generation logic is typically a simple on / off output; that is, an omnidirectional audible and visual alarm is triggered immediately when the risk indicator exceeds the threshold. In complex highway weaving zones, vehicles frequently change lanes and accelerate / decelerate. Existing technologies only employ passive observation and kinematic extrapolation for these interactions, failing to incorporate behavioral influence mechanisms between vehicles during trajectory prediction. Furthermore, the alarm strategy does not consider the risk evolution process or spatial orientation attributes.

[0004] The aforementioned existing technologies suffer from a core technical problem: collision prediction lag and alarm failure in high-interaction scenarios involving multiple vehicles on highways. Because existing technologies only extrapolate the trajectory of a single vehicle, they fail to incorporate the game-theoretic influence of surrounding vehicles' micro-behavioral interactions on trajectory evolution into the prediction model. When vehicles engage in high-interaction behaviors such as lane changes in weaving zones, the trajectory prediction results based on single-vehicle kinematics assumptions deviate significantly from the actual vehicle's trajectory. This prediction distortion directly leads to the collision time and conflict distance calculated based on the predicted trajectory failing to accurately reflect the actual hazard-causing process, thus causing a lag in risk assessment. Simultaneously, the lack of spatial orientation attributes in alarm signals prevents targeted intervention for specific risky vehicles, exacerbating alarm failure in high-interaction scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting and warning multi-source vehicle collision risks on highways, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Methods for predicting and alerting multi-source vehicle collision risks on highways include: Receive multi-vehicle trajectory time-series data generated by the fusion of multi-source sensing devices on the roadside; A dynamic spatiotemporal topology graph is constructed based on the multi-vehicle trajectory time series data, with vehicles as nodes and vehicle spacing and relative speed as edge weights. Spatiotemporal graph convolutional network is used to extract the spatiotemporal interaction features of vehicle groups in the dynamic spatiotemporal topology graph. The spatiotemporal interaction features of vehicle groups are then input into a non-cooperative game model to calculate the game equilibrium trajectory of each vehicle in the strategy space. Calculate the spatiotemporal conflict degree of the multi-vehicle game trajectory in the prediction time domain, and generate the collision risk entropy based on the time integral of the spatiotemporal conflict degree; When the collision risk entropy exceeds the dynamic adaptive threshold, a graded alarm command carrying intervention priority and spatial orientation information is generated and output to the alarm device.

[0007] Preferably, the step of constructing a dynamic spatiotemporal topology graph based on the multi-vehicle trajectory time-series data, using vehicles as nodes and vehicle spacing and relative speed as edge weights, includes: The time-series data of the multi-vehicle trajectory is sliced ​​along a time sliding window to obtain the slice data at the current moment; Map each vehicle in the current time slice data to a node feature vector; Spatial affinity is calculated based on the reciprocal of the lateral distance, longitudinal distance and relative speed between vehicles, and the spatial affinity is used as the edge weight to construct an adjacency matrix; When the spatial affinity is lower than the preset affinity threshold, the edge connection between the corresponding nodes is disconnected, and when a new vehicle enters the perception range, the corresponding node and edge connection are added to form a dynamic spatiotemporal topology graph structure that evolves with time steps.

[0008] Preferably, the step of extracting the spatiotemporal interaction features of the vehicle group in the dynamic spatiotemporal topology graph using a spatiotemporal graph convolutional network includes: In the spatial dimension, graph convolution operations are used to aggregate the node feature vectors in the dynamic spatiotemporal topology graph along the adjacency matrix to obtain the spatial structure interaction features. In the time dimension, a temporal convolution kernel is used to perform convolution operations on the spatial structure interaction features of multiple consecutive time steps along the time axis to capture the temporal evolution pattern of vehicle interaction behavior. The spatial dimension aggregation result and the temporal dimension convolution operation result are combined with feature concatenation and nonlinear mapping to output the spatiotemporal interaction features of the vehicle group that contain the spatiotemporal coupling relationship of the vehicle group.

[0009] Preferably, the step of inputting the spatiotemporal interaction characteristics of the vehicle group into a non-cooperative game model to calculate the game equilibrium trajectory of each vehicle in the strategy space includes: The revenue function for each vehicle is constructed based on the spatiotemporal interaction characteristics of the vehicle group. The revenue function includes a trajectory smoothing term and a collision penalty term. Candidate actions of each vehicle are sampled in a discrete policy space, and an optimal response policy set is constructed based on the historical policies of other vehicles. By iteratively solving for the Nash equilibrium point of each vehicle under the aforementioned payoff function, the acceleration and angular velocity decision vectors of each vehicle in the game equilibrium state are obtained. Based on the acceleration and angular velocity decision vectors, the current vehicle trajectory state is forward- deduced to generate the game equilibrium trajectory.

[0010] Preferably, the step of calculating the spatiotemporal conflict degree of the multi-vehicle game trajectory in the prediction time domain, and generating the collision risk entropy based on the time integral of the spatiotemporal conflict degree, includes: Discretize the multi-vehicle game trajectory in the prediction time domain and calculate the bounding box overlap area of ​​any two vehicles at the same time as the spatial conflict component. Calculate the time difference between any two vehicles arriving at the overlapping area of ​​the bounding box as the time conflict component. The product of the spatial conflict component and the temporal conflict component is determined as the spatiotemporal conflict degree; The spatiotemporal conflict degree is integrated along the predicted time domain, and the integration result is mapped to the non-negative real number domain to output the collision risk entropy, which represents the cumulative effect of collision evolution.

[0011] Preferably, the step of constructing the revenue function for each vehicle based on the spatiotemporal interaction characteristics of the vehicle group includes: Extract the lane departure trend vector and longitudinal acceleration vector from the spatiotemporal interaction features of the vehicle group; A lateral offset cost term is constructed based on the lane departure trend vector, and a longitudinal discomfort cost term is constructed based on the longitudinal acceleration vector; Based on the intersection points of the predicted trajectories of the current vehicle and its neighboring vehicles during the game, a dynamic collision penalty term is constructed. The lateral offset cost, longitudinal ill-adapted cost, and dynamic collision penalty are weighted and summed, with the weight coefficient of the dynamic collision penalty dynamically increasing as the spatial approximation of the predicted trajectory intersection points increases, thus generating the revenue function.

[0012] Preferably, before triggering the generation of a graded alarm command when the collision risk entropy exceeds the dynamic adaptive threshold, the step of constructing the dynamic adaptive threshold is further included: Obtain environmental visibility data and road adhesion coefficient for the current road section; A visual perception delay factor is determined based on the environmental visibility data, and a braking distance correction factor is determined based on the road adhesion coefficient. Logarithmic correction calculation is performed on the basic risk threshold based on the visual perception delay factor and braking distance correction factor; The modified base risk threshold is weighted and fused with the historical moving average of the collision risk entropy to output the dynamic adaptive threshold that changes dynamically with the environmental state and historical risk situation.

[0013] Preferably, the step of mapping each vehicle in the current time slice data to a node feature vector includes: Extract the position coordinates, velocity vectors, and heading angles of each vehicle from the slice data at the current moment; When a target occlusion loss is detected by a multi-source sensing device, a feature transfer matrix is ​​constructed based on the node feature vector of the lost target at a historical moment and the neighbor node features output by the spatiotemporal graph convolutional network. The hidden state of the occluded neighbor node is interpolated and completed using the feature transfer matrix. The interpolated features are then concatenated and dimension-reduced with the position coordinates, velocity vector, and heading angle of the currently unobstructed vehicle to generate the node feature vector after integrity restoration.

[0014] Preferably, the step of triggering the generation of a graded alarm command carrying intervention priority and spatial orientation information and outputting it to the alarm device includes: The intervention priority is generated based on the over-threshold magnitude mapping of the collision risk entropy; Based on the projection coordinates of the game equilibrium trajectory that generates the collision risk entropy onto the two-dimensional plane, the yaw angle and pitch angle of the alarm beam are calculated as the spatial orientation information. The intervention priority and the spatial orientation information are encapsulated into a control signaling frame and sent to the roadside phased array alarm device. The roadside phased array alarm device analyzes the control signaling frame and focuses and projects the audible and visual alarm beam onto the spatial area where the corresponding risk vehicle is located based on the yaw angle and pitch angle.

[0015] Preferably, before the step of receiving the multi-vehicle trajectory time-series data fused from the roadside multi-source sensing devices, the method further includes a step of time alignment of the multi-source sensing data: Receive initial radar point cloud data output from roadside millimeter-wave radar and initial image frame data output from visual sensor; Extract the radar timestamp from the initial radar point cloud data and the image timestamp from the initial image frame data; Calculate the time difference between radar timestamp and image timestamp, and use the data source with the earlier timestamp as a reference to perform polynomial fitting interpolation on the trajectory of the moving target in the data source with the later timestamp. The interpolated moving target trajectory is spatially registered and fused with the data from the reference data source to generate time-synchronized multi-vehicle trajectory time-series data.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention solves the technical problems of collision prediction lag and alarm failure in multi-vehicle strong interaction scenarios by constructing a dynamic spatiotemporal topology graph and combining it with a spatiotemporal graph convolutional network and a non-cooperative game model. A dynamic spatiotemporal topology graph is constructed with vehicles as nodes and vehicle spacing and relative speed as edge weights, mapping isolated vehicle trajectories into a physically related graph structure. The spatiotemporal graph convolutional network extracts the spatiotemporal interaction features of the vehicle group along this graph structure, capturing the spatial position dependence and temporal evolution patterns among multiple vehicles. The spatiotemporal interaction features are input into the non-cooperative game model to calculate the game equilibrium trajectory. A strategy influence mechanism between vehicles is introduced in the trajectory prediction stage, replacing the traditional single-vehicle independent kinematic extrapolation, making the predicted trajectory closely match the actual driving intentions during multi-vehicle strong interaction. Based on the multi-vehicle game trajectory, the spatiotemporal conflict degree is calculated and time-integrated to generate collision risk entropy. This quantifies and combines transient spatial conflict with temporal risk accumulation. When the collision risk entropy exceeds a dynamic adaptive threshold, a tiered alarm command is triggered. This mechanism overcomes the discontinuity of geometric threshold determination at a single moment, realizes continuous measurement of the collision risk evolution process, eliminates the prediction blind spot of strong interaction-induced risk process from a mechanistic perspective, and ensures the timeliness of risk determination.

[0017] 2. This invention further optimizes the accuracy of risk prediction and the reliability of alarms through multiple technical means. In the payoff function of the game model, a cost term is constructed based on lane departure trends and longitudinal acceleration, and a collision penalty term weight that dynamically increases with the degree of spatial approximation is set, making the solution process of the game equilibrium trajectory conform to driving physical constraints and risk avoidance logic. Visual perception delay factors and braking distance correction factors are determined based on environmental visibility and road adhesion coefficients. The basic risk threshold is logarithmically corrected and fused with the historical moving average to generate a dynamic adaptive threshold, allowing the alarm trigger boundary to dynamically adjust with the external physical environment, avoiding false alarms under extreme weather conditions caused by fixed thresholds. When target occlusion occurs in multi-source sensing devices, the hidden state of the occluded neighboring nodes is interpolated and completed using a feature transfer matrix, maintaining the integrity of the dynamic spatiotemporal topology. In the data input stage, by calculating the time difference between radar timestamps and image timestamps, polynomial fitting interpolation and spatial coordinate registration are performed on the time-lagging data source, eliminating the time asynchrony bias of multi-source data. When a graded alarm command is triggered, the yaw and pitch angles are calculated based on the projection coordinates of the equilibrium trajectory of the game. The phased array alarm device is then driven to focus and project the audible and visual alarm beam onto the space area where the at-risk vehicle is located. This achieves directional spatial projection of the alarm signal and avoids interference from omnidirectional alarms to unrelated lanes. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the overall process of the multi-source vehicle collision risk prediction and alarm method for highways according to the present invention. Figure 2 This is a flowchart of the dynamic spatiotemporal topology graph construction and updating process of the present invention; Figure 3 This is a flowchart of the spatiotemporal graph convolutional network feature extraction process of the present invention; Figure 4 The flowchart for solving the equilibrium trajectory in the non-cooperative game model of this invention is shown below; Figure 5 This is a flowchart of the collision risk entropy calculation and graded alarm determination process of the present invention; Figure 6 This is a flowchart of the multi-source sensing data time alignment and fusion preprocessing process of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please refer to Figure 1This embodiment provides a method for predicting and alarming multi-source vehicle collision risks on highways. By constructing a dynamic spatiotemporal topology map and combining a spatiotemporal graph convolutional network with a non-cooperative game model, it achieves accurate prediction and targeted alarm of collision risks in multi-vehicle strong interaction scenarios on highways. The overall workflow of this invention is as follows: First, it receives multi-vehicle trajectory time-series data generated by roadside multi-source sensing devices and constructs a dynamic spatiotemporal topology map with vehicles as nodes and vehicle spacing and relative speed as edge weights. Then, it uses a spatiotemporal graph convolutional network to extract the spatiotemporal interaction features of the vehicle group in the dynamic spatiotemporal topology map, and inputs the spatiotemporal interaction features of the vehicle group into a non-cooperative game model to calculate the game equilibrium trajectory of each vehicle in the strategy space. Next, it calculates the spatiotemporal conflict degree of the multi-vehicle game trajectory in the prediction time domain, and generates a collision risk entropy based on the time integral of the spatiotemporal conflict degree. Finally, when the collision risk entropy exceeds the dynamic adaptive threshold, it triggers the generation of a hierarchical alarm command carrying intervention priority and spatial orientation information and outputs it to the alarm device.

[0021] In one embodiment, the roadside multi-source sensing equipment includes millimeter-wave radar and visual sensors, deployed on poles on both sides of the highway. The coverage areas of adjacent sensing devices overlap to ensure continuous sensing of the entire highway section. The millimeter-wave radar acquires vehicle distance, speed, and angle information, while the visual sensors acquire vehicle image information and lane line information. The multi-source sensing equipment transmits the collected raw data to a roadside edge computing unit, which performs data fusion, trajectory generation, risk prediction, and alarm command generation. The edge computing unit interacts with the highway management center's cloud platform via a fiber optic network and simultaneously transmits control signals to the roadside phased array alarm device via a wireless communication module.

[0022] In one embodiment, multi-vehicle trajectory time-series data generated by roadside multi-source sensing devices is received. The multi-vehicle trajectory time-series data includes the position coordinates, velocity vectors, heading angles, and vehicle type information of multiple vehicles within consecutive time steps. The trajectory data of each vehicle is stored indexed by a timestamp, forming a time-series structure. The sampling frequency of the multi-vehicle trajectory time-series data is consistent with the sampling frequency of the roadside multi-source sensing devices to ensure the temporal continuity of the trajectory data.

[0023] Using vehicles as nodes and vehicle spacing and relative speed as edge weights, a dynamic spatiotemporal topology graph is constructed based on multi-vehicle trajectory time-series data. The dynamic spatiotemporal topology graph is a graph structure that evolves over time, where each node represents a vehicle and each edge represents the spatial interaction between two vehicles. The edge weights are determined by the vehicle spacing and relative speed, reflecting the intensity of the interaction between the two vehicles. When a vehicle enters the sensing range, a corresponding node and its edge connection to other nodes are added to the graph; when a vehicle leaves the sensing range, the corresponding node and its edge connection to other nodes are deleted; when the interaction intensity between vehicles falls below a preset threshold, the edge connections between corresponding nodes are disconnected.

[0024] A spatiotemporal graph convolutional network (SPGraph Convolutional Network) is employed to extract spatiotemporal interaction features of vehicle groups in a dynamic spatiotemporal topology graph. The SPGraph Convolutional Network extracts features simultaneously in both spatial and temporal dimensions. In the spatial dimension, graph convolution operations are used to capture the spatial positional dependencies between vehicles, while in the temporal dimension, temporal convolution operations are used to capture the temporal evolution of vehicle interaction behaviors. The SPGraph Convolutional Network takes node feature vectors and adjacency matrices from the dynamic spatiotemporal topology graph as input and outputs spatiotemporal interaction features of the vehicle group that include the spatiotemporal coupling relationships within the vehicle group.

[0025] The spatiotemporal interaction characteristics of the vehicle group are input into a non-cooperative game model to calculate the game equilibrium trajectory of each vehicle in the policy space. The non-cooperative game model treats each vehicle as an independent player, and each player selects acceleration and angular velocity decision vectors in a discrete policy space to maximize its own payoff function. The payoff function includes a trajectory smoothing term and a collision penalty term, which are used to constrain the vehicle's driving comfort and safety, respectively. By iteratively solving for the Nash equilibrium point of each vehicle under the payoff function, the decision vector of each vehicle in the game equilibrium state is obtained. Based on this decision vector, the current vehicle trajectory state is forward extrapolated to generate the game equilibrium trajectory.

[0026] The spatiotemporal conflict degree of the multi-vehicle game trajectory is calculated in the prediction time domain, and the collision risk entropy is generated based on the time integral of the spatiotemporal conflict degree. The prediction time domain is a future time interval, the length of which is determined by the highway's design speed and vehicle braking distance. The multi-vehicle game trajectory is discretized and sampled within the prediction time domain. The overlapping area of ​​the bounding boxes of any two vehicles at the same time is calculated as the spatial conflict component, and the time difference between the arrival times of any two vehicles at the overlapping area is calculated as the temporal conflict component. The product of the spatial and temporal conflict components is determined as the spatiotemporal conflict degree. The spatiotemporal conflict degree is integrated along the prediction time domain, and the integral result is mapped to the non-negative real number domain to output the collision risk entropy, which characterizes the cumulative effect of collision evolution.

[0027] When the collision risk entropy exceeds the dynamic adaptive threshold, a tiered alarm command carrying intervention priority and spatial orientation information is triggered and output to the alarm device. The dynamic adaptive threshold is dynamically adjusted based on the current road segment's environmental visibility data and road adhesion coefficient, and is also corrected by combining the historical moving average of the collision risk entropy. The intervention priority is generated based on the magnitude of the collision risk entropy exceeding the threshold; the greater the magnitude of the exceedance, the higher the intervention priority. The spatial orientation information is calculated based on the projection coordinates of the game equilibrium trajectory that generates the collision risk entropy onto a two-dimensional plane, including the yaw angle and pitch angle of the alarm beam. The tiered alarm command is encapsulated into a control signaling frame and sent to the roadside phased array alarm device. After parsing the control signaling frame, the roadside phased array alarm device focuses and projects the audible and visual alarm beam onto the spatial area where the corresponding risk vehicle is located, based on the yaw angle and pitch angle.

[0028] In this embodiment, isolated vehicle trajectories are mapped to physically related graph structures by constructing a dynamic spatiotemporal topology graph, enabling explicit modeling of spatial interactions between vehicles. A spatiotemporal graph convolutional network is used to extract spatiotemporal interaction features of the vehicle group, simultaneously capturing the spatial positional dependencies and temporal evolution patterns among multiple vehicles. These spatiotemporal interaction features are input into a non-cooperative game model to calculate the equilibrium trajectory. A strategy influence mechanism between vehicles is introduced during the trajectory prediction stage, ensuring that the predicted trajectory closely matches the actual driving intentions of multiple vehicles during periods of strong interaction. Based on the multi-vehicle game trajectory, the spatiotemporal conflict degree is calculated and time-integrated to generate a collision risk entropy. This quantifies and combines transient spatial conflict with temporal risk accumulation, achieving continuous measurement of the collision risk evolution process. When the collision risk entropy exceeds a dynamic adaptive threshold, a tiered alarm command is triggered, driving the phased array alarm device to focus and project the audible and visual alarm beam onto the spatial region where the at-risk vehicle is located, achieving directional spatial projection of the alarm signal.

[0029] In one embodiment, reference Figure 2 The system slices the time-series trajectory data of multiple vehicles along a sliding time window to obtain the current time-stack slice data. The length of the sliding time window is a fixed value, and its size is determined according to the time scale of vehicle interaction behavior. The sliding time window slides along the time axis in fixed steps, generating a new current time-stack slice data with each slide. The current time-stack slice data contains the trajectory data of all vehicles within the sliding time window, and the trajectory data of each vehicle includes the position coordinates, velocity vector, and heading angle information of multiple time steps within the window.

[0030] Each vehicle in the current time-slice of data is mapped to a node feature vector. The node feature vector is a quantified representation of the vehicle's current state, containing its position coordinates, velocity vector, heading angle, and vehicle type information. For each vehicle, its latest state information in the current time-slice of data is extracted and concatenated to form the initial node feature vector. The dimension of the initial node feature vector is determined based on the amount of state information it contains.

[0031] Spatial affinity is calculated based on the reciprocal of the lateral and longitudinal distances and relative speeds between vehicles. This spatial affinity is then used as edge weights to construct an adjacency matrix. Spatial affinity quantifies the spatial interaction strength between two vehicles, and its calculation formula is as follows: , in, Indicates vehicle With vehicles Spatial affinity between them Indicates vehicle With vehicles Horizontal distance between them Indicates vehicle With vehicles The longitudinal distance between them Indicates vehicle With vehicles The relative speed between them , , These are the weighting coefficients. To prevent small constants with a denominator of zero, weighting coefficients are used. , , Based on the traffic flow characteristics of highways, these parameters are predetermined and used to adjust the degree of influence of lateral distance, longitudinal distance, and relative speed on spatial affinity.

[0032] Adjacency matrix It is The square array, in which This represents the number of vehicles in the slice of data at the current time. Elements of the adjacency matrix. Indicates vehicle With vehicles The edge weights between them, when hour, This means that the vehicle itself has no edge connections. The adjacency matrix is ​​a symmetric matrix, i.e. This indicates that the interaction between vehicles is mutual.

[0033] When the spatial affinity is lower than a preset affinity threshold, the edge connections between corresponding nodes are disconnected, and when a new vehicle enters the perception range, the corresponding nodes and edge connections are added, forming a dynamic spatiotemporal topology graph structure that evolves over time. The preset affinity threshold is used to filter vehicle pairs with weak interaction strength, reducing the complexity and computational cost of the graph structure. When a vehicle... With vehicles Spatial affinity between When the spatial affinity falls below a preset affinity threshold, the corresponding element in the adjacency matrix is ​​set to zero, indicating that the edge connection between the two nodes is broken. When a new vehicle enters the perception range, a row and a column are added to the adjacency matrix, the spatial affinity between the new vehicle and all existing vehicles is calculated, and the new row and column are filled. When a vehicle leaves the perception range, the corresponding row and column are deleted from the adjacency matrix.

[0034] In this embodiment, the continuous time-series data of multi-vehicle trajectories is sliced ​​along a sliding time window, converting the continuous time-series data into discrete time-step data, which facilitates subsequent graph structure construction and feature extraction. Spatial affinity is calculated based on the reciprocal of the lateral distance, longitudinal distance, and relative speed between vehicles, ensuring that edge weights accurately reflect the spatial interaction strength between vehicles. By setting a preset affinity threshold to filter vehicle pairs with weak interaction strength, computational complexity is reduced while ensuring that the graph structure can capture the main interaction relationships. The nodes and edge connections in the graph structure are dynamically updated, enabling the dynamic spatiotemporal topology graph to reflect the changes in highway traffic flow in real time.

[0035] In a preferred embodiment, the position coordinates, velocity vector, and heading angle of each vehicle in the current time slice data are extracted. The position coordinates are represented in a geodetic coordinate system, including longitude, latitude, and altitude information. The velocity vector is a two-dimensional vector containing lateral and longitudinal velocity components. The heading angle is the angle between the vehicle's direction of travel and true north, and its value ranges from [value missing]. .

[0036] When target occlusion loss is detected by multi-source sensing devices, a feature transfer matrix is ​​constructed based on the node feature vectors of the lost target at historical time points and the neighbor node features output by the spatiotemporal graph convolutional network. Target occlusion loss refers to the inability of roadside multi-source sensing devices to detect the presence of a certain vehicle due to obstacle occlusion or performance limitations of the sensing devices. When target occlusion loss is detected, the identifier and time of loss of the lost target are recorded, and the node feature vectors of the lost target at multiple consecutive historical time points before the time of loss are obtained. Simultaneously, the node features of nodes adjacent to the lost target output by the spatiotemporal graph convolutional network are obtained; adjacent nodes are those connected to the lost target by edges before the time of loss.

[0037] The feature transfer matrix describes the process of transferring features from neighboring nodes to the lost target node, and its calculation formula is as follows: , in, Represents the feature transfer matrix, For a degree matrix, its diagonal elements , This is the submatrix in the adjacency matrix related to the lost target at the time of loss. Each row of the feature transfer matrix corresponds to a neighbor node, and each column corresponds to a feature dimension of the lost target node.

[0038] The hidden states of occluded neighbor nodes are interpolated and completed using the feature transfer matrix. The hidden states of occluded neighbor nodes are obtained by multiplying the neighbor node features by the feature transfer matrix; the calculation formula is as follows: ,

[0039] in, This represents the hidden state vector of the occluded neighbor node. This represents the feature matrix of neighboring nodes, where each row corresponds to the feature vector of a neighboring node. The hidden state vector obtained through interpolation completion contains estimates of the position, velocity, and heading angle of the lost target node.

[0040] The interpolated and completed features are concatenated and dimension-reduced with the position coordinates, velocity vector, and heading angle of the currently unoccluded vehicle to generate the node feature vector after integrity restoration. For unoccluded vehicles, their current position coordinates, velocity vector, and heading angle are directly extracted as feature vectors. For occluded vehicles, the hidden state vector obtained through interpolation is used as their feature vector. The feature vectors of all vehicles are concatenated to form a high-dimensional feature matrix, and then a linear transformation is used to map the high-dimensional feature matrix to a fixed-dimensional feature space to generate the node feature vector matrix after integrity restoration.

[0041] In this embodiment, when a target is lost due to occlusion in the multi-source sensing device, a feature transfer matrix is ​​constructed using the node feature vector of the lost target at a historical moment and the features of neighboring nodes. This matrix is ​​then used to interpolate and complete the hidden state of the occluded neighboring nodes, maintaining the integrity of the dynamic spatiotemporal topology. The feature transfer matrix enables the transfer of neighboring node features to the lost target node, allowing the completed features to reflect the interaction between the lost target node and surrounding vehicles. The interpolated and completed features are then concatenated and dimension-reduced with the features of unoccluded vehicles to generate a node feature vector matrix of uniform dimension, facilitating subsequent spatiotemporal graph convolutional network processing.

[0042] In one embodiment, reference Figure 3In the spatial dimension, graph convolution operations are used to aggregate the feature vectors of nodes in a dynamic spatiotemporal topology graph along the adjacency matrix to obtain spatial structural interaction features. Graph convolution operations capture spatial dependencies in the graph structure by weighted aggregation of a node's own features with the features of its neighboring nodes. The calculation formula for graph convolution operations is: , in, Indicates the first The input feature matrix of layer graph convolution, Indicates the first The output feature matrix of layer graph convolution, To add a self-loop adjacency matrix, It is the identity matrix. The diagonal elements of the degree matrix after adding self-loops. , For the first The learnable weight matrix of layer graph convolution. It is a non-linear activation function.

[0043] In the spatial dimension, multiple layers of graph convolutional operations are stacked, enabling the features of each node to aggregate information from its multi-hop neighbors. The first layer of graph convolution aggregates the features of each node's first-order neighbors, the second layer aggregates the features of each node's second-order neighbors, and so on. Through these multiple layers of graph convolutional operations, node features can encompass a wider range of spatial interaction information. The output of the last layer of graph convolution is used as the spatial structure interaction feature.

[0044] In the time dimension, a temporal convolution kernel is used to perform convolution operations along the time axis on the spatial structural interaction features of multiple consecutive time steps to capture the temporal evolution of vehicle interaction behavior. The temporal convolution operation uses a one-dimensional convolution kernel to perform sliding convolution on the feature sequence in the time dimension. The size of the temporal convolution kernel determines the length of the temporal dependencies that can be captured. The calculation formula for the temporal convolution operation is: , in, Indicates time Temporal convolution output features Indicates time Spatial structural interaction characteristics, Indicates the temporal convolution kernel at the offset The weight matrix at the location, The size of the temporal convolution kernel. For bias vectors, This represents the convolution operation.

[0045] In the time dimension, multiple layers of temporal convolution operations are stacked to capture the temporal evolution patterns at different time scales. Different kernel sizes are used in different layers of temporal convolution to capture the interactive changes between short and long time scales. The output of the last layer of temporal convolution is used as the temporal structure interaction feature.

[0046] The spatial aggregation result and the temporal convolution result are concatenated and nonlinearly mapped to output the spatiotemporal interaction features of the vehicle group, which contain the spatiotemporal coupling relationship of the vehicle group. The spatial structure interaction features and the temporal structure interaction features are concatenated along the feature dimension to form a high-dimensional spatiotemporal feature vector. Then, a fully connected layer is used to nonlinearly map the concatenated spatiotemporal feature vector to a fixed-dimensional feature space, generating the spatiotemporal interaction features of the vehicle group. Each dimension of the spatiotemporal interaction features of the vehicle group corresponds to a quantized representation of a vehicle group interaction mode.

[0047] In this embodiment, in the spatial dimension, graph convolution operations are used to aggregate node features along the adjacency matrix, capturing the spatial positional dependencies between vehicles. In the temporal dimension, a temporal convolution kernel is used to perform convolution operations along the time axis, capturing the temporal evolution of vehicle interaction behavior. The features from the spatial and temporal dimensions are concatenated and nonlinearly mapped to generate spatiotemporal interaction features of the vehicle group that include the spatiotemporal coupling relationships of the vehicle group, enabling the features to simultaneously reflect the interaction characteristics of the vehicle group in both space and time.

[0048] In one embodiment, reference Figure 4 Based on the spatiotemporal interaction characteristics of the vehicle group, a benefit function for each vehicle is constructed. This function includes a trajectory smoothing term and a collision penalty term. The benefit function quantifies the gain a vehicle receives when choosing a particular decision vector; the vehicle's goal is to choose a decision vector that maximizes its own benefit function. The trajectory smoothing term constrains the vehicle's driving comfort, preventing abrupt acceleration, deceleration, or steering maneuvers. The collision penalty term constrains the vehicle's driving safety, preventing collisions with other vehicles.

[0049] Candidate actions for each vehicle are sampled in a discrete policy space, and an optimal response policy set is constructed based on the historical policies of other vehicles. The policy space consists of the vehicle's acceleration and angular velocity decision vectors, with the acceleration value ranging from... The range of angular velocity is The ranges of acceleration and angular velocity are discretized into multiple equally spaced values, forming a discrete policy space. For each vehicle, multiple candidate actions are sampled from the discrete policy space as its possible decision vectors.

[0050] For each candidate action based on the historical strategies of other vehicles, calculate the payoff function value. Select the candidate action that maximizes the payoff function value as the optimal response for that vehicle to the historical strategies of other vehicles. Combine the optimal responses of all vehicles to form the optimal response strategy set. The optimal response strategy set represents the optimal decision vector for each vehicle when other vehicles adopt their historical strategies.

[0051] By iteratively solving for the Nash equilibrium point of each vehicle under the payoff function, the acceleration and angular velocity decision vectors of each vehicle in the game equilibrium state are obtained. A Nash equilibrium point is a strategy combination under which no vehicle can improve its payoff function value by unilaterally changing its own strategy. The iterative solution process is as follows: First, initialize the strategies of each vehicle as randomly selected candidate actions; then, in each iteration, each vehicle calculates its optimal response strategy based on the strategies of other vehicles in the previous iteration; then update the strategies of each vehicle to the optimal response strategy; repeat the above process until the strategies of each vehicle converge or the maximum number of iterations is reached. When the strategies converge, the current strategy combination is the Nash equilibrium point.

[0052] Based on the acceleration and angular velocity decision vectors, a forward inference is performed on the current vehicle trajectory state to generate a game equilibrium trajectory. The current vehicle trajectory state includes the vehicle's position coordinates, velocity vector, and heading angle. According to the vehicle kinematics model, the acceleration and angular velocity decision vectors are used to forward infer the current vehicle trajectory state, obtaining the vehicle's trajectory state at multiple future time steps. The calculation formula for the vehicle kinematics model is as follows: , in, Indicates the vehicle's time Location coordinates, Indicates the vehicle's time speed magnitude, Indicates the vehicle's time The heading angle, Indicates the vehicle's time acceleration, Indicates the vehicle's time angular velocity, Indicates the time step.

[0053] By connecting the vehicle's trajectory states at multiple future time steps, a game equilibrium trajectory is formed. The game equilibrium trajectory contains the vehicle's position coordinates, velocity vector, and heading angle information at each time step within the prediction time domain.

[0054] In this embodiment, each vehicle is treated as an independent player in the game, and the driving comfort and safety of the vehicle are quantified by constructing a payoff function. Candidate actions of each vehicle are sampled in a discrete policy space, and an optimal response policy set is constructed based on the historical policies of other vehicles. By iteratively solving for the Nash equilibrium point, the decision vector of each vehicle in the game equilibrium state is obtained, enabling the predicted trajectory to reflect the policy interaction between vehicles. Based on the vehicle kinematic model, the current vehicle trajectory state is forward extrapolated to generate the game equilibrium trajectory, providing a foundation for subsequent collision risk calculation.

[0055] In a preferred embodiment, a lane departure trend vector and a longitudinal acceleration vector are extracted from the spatiotemporal interaction features of the vehicle group. The lane departure trend vector represents the tendency of a vehicle to deviate from the centerline of the current lane; its magnitude and direction reflect the probability and direction of lane changing. The longitudinal acceleration vector represents the rate of change of the vehicle's longitudinal acceleration; its magnitude reflects the severity of acceleration and deceleration. The lane departure trend vector and the longitudinal acceleration vector are obtained by performing a linear transformation on the spatiotemporal interaction features of the vehicle group.

[0056] A lateral offset cost term is constructed based on the lane departure trend vector, and a longitudinal ill-adjusted cost term is constructed based on the longitudinal acceleration vector. The lateral offset cost term quantifies the cost of a vehicle deviating from the current lane centerline, and its calculation formula is as follows: , in, This represents the lateral offset cost term. Represents the lane departure trend vector. This represents the weighting coefficient of the lateral offset cost term. The lateral offset cost term is proportional to the square of the magnitude of the lane offset trend vector; the greater the tendency of the vehicle to deviate from the lane centerline, the larger the value of the lateral offset cost term.

[0057] The longitudinal discomfort cost term is used to quantify the cost resulting from the severity of longitudinal acceleration and deceleration of the vehicle. Its calculation formula is as follows: , in, This represents the vertically unsuitable cost term. Represents the longitudinal acceleration vector. This represents the weighting coefficient of the longitudinal adverse cost term. The longitudinal adverse cost term is proportional to the square of the magnitude of the longitudinal acceleration vector; the more severe the vehicle's acceleration or deceleration, the larger the value of the longitudinal adverse cost term.

[0058] A dynamic collision penalty term is constructed based on the intersection points of the predicted trajectories of the current vehicle and its neighboring vehicles during the game. The intersection point is the point on a two-dimensional plane where the predicted trajectories of the current vehicle and its neighboring vehicle intersect. When an intersection point exists, the time difference and distance difference between the current vehicle and its neighboring vehicle reaching the intersection point are calculated. The formula for calculating the dynamic collision penalty term is: , in, Indicates the dynamic collision penalty. This indicates the time difference between the current vehicle and the adjacent vehicle arriving at the intersection. This indicates the distance between the current vehicle and its adjacent vehicle at the intersection. and For scale parameters, This represents the dynamic weighting coefficient, whose value varies with distance. The dynamic weighting coefficient increases as the value decreases. The formula for calculating the dynamic weighting coefficient is: , in, Indicates the basic weighting coefficient. Indicates the reference distance. This represents the attenuation coefficient. When the distance between vehicles is less than the reference distance, the dynamic weighting coefficient increases exponentially with decreasing distance, causing the collision penalty to increase rapidly as vehicles approach each other.

[0059] The lateral offset cost, longitudinal ill-adjusted cost, and dynamic collision penalty are weighted and summed to generate the profit function. The formula for calculating the profit function is: , in, Let represent the payoff function. Since a larger cost term results in a lower vehicle payoff, the payoff function is the negative of the sum of the cost terms. The vehicle's goal is to choose a decision vector that maximizes the payoff function, i.e., minimizes the sum of the cost terms.

[0060] In this embodiment, lateral offset cost and longitudinal discomfort cost terms are constructed based on the lane departure trend vector and the longitudinal acceleration vector, enabling the payout function to reflect the vehicle's driving comfort constraints. A dynamic collision penalty term is constructed based on the intersection points of the predicted trajectories of the current vehicle and adjacent vehicles, with a weighting coefficient that dynamically increases with spatial proximity, ensuring the payout function reflects the vehicle's driving safety constraints, with the weight of safety constraints rapidly increasing as vehicles approach each other. The payout function is generated by weighted summation of the lateral offset cost term, the longitudinal discomfort cost term, and the dynamic collision penalty term, ensuring that the solution process for the game equilibrium trajectory aligns with driving physics constraints and risk avoidance logic.

[0061] In one embodiment, reference Figure 5In the prediction time domain, the multi-vehicle game trajectories are discretized and sampled. The overlapping area of ​​the bounding boxes of any two vehicles at the same time is calculated as the spatial conflict component. The length of the prediction time domain is... Discretize the prediction time domain into There are 1 equally spaced time steps, each time step having a length of 1. For each time step, extract the position coordinates and heading angles of all vehicles at that time step, and construct the vehicle bounding box. The vehicle bounding box is rectangular, and its length and width are determined according to the vehicle type.

[0062] For any two cars and Calculate their time at the same moment The bounding box overlap area is calculated as follows: First, convert the two rectangular bounding boxes to vertex coordinates in the same coordinate system. Then, calculate the intersection area of ​​the two rectangles, and finally calculate the area of ​​the intersection area. If the two bounding boxes do not intersect, the overlap area is zero. The bounding box overlap area is used as the vehicle's... With vehicles At any moment Spatial conflict components, denoted as .

[0063] Calculate the time difference between any two vehicles reaching the overlapping area of ​​their bounding boxes as the time conflict component. When vehicles With vehicles The enclosing box at all times When overlap exists, calculate vehicles Time to reach the overlapping region With vehicles Time to reach the overlapping region The time to reach the overlapping area refers to the time required for a vehicle to travel from its current position to the boundary of the overlapping area. The time conflict component is the absolute value of the time difference between the arrival times of the two vehicles in the overlapping area, denoted as . If the two bounding boxes do not intersect, the time conflict component is infinite.

[0064] The product of the spatial conflict component and the temporal conflict component is defined as the spatiotemporal conflict degree. The spatiotemporal conflict degree is used to quantify the time-space conflict between two vehicles. The degree of collision risk is calculated using the following formula: , in, Indicates vehicle With vehicles At any moment The degree of spatiotemporal conflict, To prevent small constants with a denominator of zero, the spatiotemporal conflict degree is directly proportional to the spatial conflict component and inversely proportional to the temporal conflict component. The larger the overlap area of ​​the enclosure boxes, the smaller the time difference between the two vehicles arriving at the overlap area, the higher the value of the spatiotemporal conflict degree, indicating a higher collision risk.

[0065] The spatiotemporal conflict degree is integrated along the prediction time domain, and the integration result is mapped to the non-negative real number domain to output the collision risk entropy, which represents the cumulative effect of collision evolution. For any two vehicles... and Integrating the spatiotemporal conflict degree along the prediction time domain yields the cumulative collision risk of the two vehicles: , Summing the cumulative collision risks of all vehicle pairs yields the total cumulative collision risk: , in, Let be the total number of vehicles. The total cumulative collision risk is mapped to a non-negative real number domain to generate the collision risk entropy. The mapping function uses a logarithmic function, and its calculation formula is: , in, This represents the collision risk entropy. The higher the collision risk entropy value, the higher the collision risk.

[0066] In this embodiment, the spatial conflict component is calculated by measuring the overlap area of ​​the bounding boxes of any two vehicles at the same time, and the temporal conflict component is calculated by measuring the time difference between the arrival times of the two vehicles at the overlap area. This combination of spatial and temporal conflict quantifies the degree of collision risk. The spatiotemporal conflict degree is integrated along the prediction time domain to obtain the cumulative effect of collision risk. The total cumulative collision risk is then mapped to the non-negative real number domain using a logarithmic function to generate the collision risk entropy, thus achieving a continuous measurement of the collision risk evolution process.

[0067] In one embodiment, environmental visibility data and road adhesion coefficient for the current road segment are acquired. Environmental visibility data is acquired through visibility sensors deployed on the roadside, and the unit is meters. The road adhesion coefficient is acquired through road surface condition sensors deployed on the roadside, and its value ranges from [value missing]. A higher value indicates better road surface adhesion performance. The sampling frequency of environmental visibility data and road adhesion coefficient is consistent with the sampling frequency of multi-vehicle trajectory time series data.

[0068] The visual perception delay factor is determined based on environmental visibility data, and the braking distance correction factor is determined based on the road adhesion coefficient. The visual perception delay factor is used to quantify the impact of environmental visibility on the driver's visual perception time, and its calculation formula is as follows: , in, Indicates the visual perception delay factor. Indicates standard visibility. This indicates the current ambient visibility. When ambient visibility is lower than the standard visibility, the visual perception delay factor is greater than 1, indicating that the driver's visual perception time is prolonged.

[0069] The braking distance correction factor is used to quantify the impact of road adhesion coefficient on vehicle braking distance, and its calculation formula is as follows: , in, This represents the braking distance correction factor. Indicates the standard road adhesion coefficient. This indicates the current road adhesion coefficient. When the road adhesion coefficient is lower than the standard road adhesion coefficient, the braking distance correction factor is greater than 1, indicating that the vehicle's braking distance increases.

[0070] A logarithmic correction is performed on the baseline risk threshold based on the visual perception delay factor and braking distance correction factor. The baseline risk threshold is the collision risk entropy threshold determined under standard environmental conditions. The formula for the logarithmic correction is as follows: , in, This represents the risk threshold after a correction. This represents the basic risk threshold. Through logarithmic correction, the risk threshold increases with the increase of the visual perception delay factor and the braking distance correction factor, thereby increasing the risk threshold and reducing false alarms under adverse environmental conditions.

[0071] The revised baseline risk threshold is weighted and fused with the historical moving average of collision risk entropy to output a dynamically adaptive threshold that changes dynamically with environmental conditions and historical risk trends. The historical moving average of collision risk entropy refers to the average value of collision risk entropy over a past period, and its calculation formula is as follows: , in, This represents the historical moving average of the collision risk entropy. Indicates the length of the sliding window. Indicates time Collision risk entropy.

[0072] The formula for weighted fusion is: , in, Indicates a dynamic adaptive threshold. This represents the weighting coefficient, with a value range of [value range missing]. The weighting coefficients are used to adjust the degree to which environmental conditions and historical risk profiles influence the dynamic adaptive threshold. When When the dynamic adaptive threshold is determined solely by the environmental state, when... At that time, the dynamic adaptive threshold is determined solely by the historical risk situation.

[0073] In this embodiment, a visual perception delay factor and a braking distance correction factor are determined based on environmental visibility data and road adhesion coefficient. The basic risk threshold is then logarithmically corrected, allowing the alarm trigger boundary to dynamically adjust with the external physical environment. The corrected basic risk threshold is weighted and fused with the historical moving average of the collision risk entropy, enabling the dynamic adaptive threshold to simultaneously reflect both the current environmental state and historical risk trends. By constructing a dynamic adaptive threshold, false alarms and missed alarms under extreme weather conditions are avoided, as with fixed thresholds.

[0074] In one embodiment, intervention priorities are generated based on the over-threshold magnitude of the collision risk entropy. The over-threshold magnitude refers to the portion of the collision risk entropy that exceeds a dynamically adaptive threshold, and its calculation formula is as follows: , in, Indicates the magnitude of exceeding the threshold. When At that time, no alarm is triggered.

[0075] Intervention priorities are divided into multiple levels, with higher levels indicating more urgent interventions. The mapping relationship between threshold amplitude and intervention priority is determined by a pre-defined piecewise function. The mapping relationship between threshold amplitude and intervention priority is shown in Table 1.

[0076] Table 1. Mapping relationship between threshold magnitude and intervention priority

[0077] In Table 1, , , Pre-set thresholds are used to classify different intervention priority levels. Different intervention priorities correspond to different alarm types: alert alarms remind the driver to pay attention to surrounding vehicles, warning alarms remind the driver of potential collision risks, emergency alarms remind the driver to take immediate evasive action, and hazard alarms remind the driver that a collision is imminent.

[0078] Based on the projected coordinates of the game equilibrium trajectory that generates collision risk entropy onto a two-dimensional plane, the yaw and pitch angles of the alarm beam are calculated as spatial orientation information. First, the vehicle pair generating the maximum spatiotemporal conflict intensity is identified, and the projected coordinates of their game equilibrium trajectories onto a two-dimensional plane are extracted. The midpoint coordinates of the two vehicles' projected coordinates are calculated as the center coordinates of the risk area. Based on the center coordinates of the risk area and the installation location coordinates of the roadside phased array alarm device, the yaw and pitch angles of the alarm beam are calculated.

[0079] Yaw angle refers to the angle between the alarm beam on the horizontal plane and true north, and its calculation formula is: , in, Indicates the yaw angle. Indicates the center coordinates of the risk area. This indicates the coordinates of the installation location of the roadside phased array alarm device.

[0080] Pitch angle refers to the angle between the alarm beam on the vertical plane and the horizontal plane, and its calculation formula is: , in, Indicates pitch angle, Indicates the center height of the risk area. This indicates the installation height of the roadside phased array alarm device.

[0081] Intervention priority and spatial orientation information are encapsulated into control signaling frames and sent to the roadside phased array alarm device. The control signaling frames use a fixed format, including a frame header, intervention priority field, yaw angle field, pitch angle field, and frame trailer. The frame header and trailer identify the start and end of the control signaling frame, the intervention priority field transmits intervention priority information, and the yaw angle and pitch angle fields transmit spatial orientation information. The control signaling frames are transmitted to the roadside phased array alarm device via a wireless communication module.

[0082] The roadside phased array alarm device analyzes control signal frames and focuses the audible and visual alarm beams onto the spatial area where the corresponding at-risk vehicle is located, based on the yaw and pitch angles. The roadside phased array alarm device includes a phased array antenna array, an audible and visual generator, and a control unit. The control unit analyzes the received control signal frames and extracts intervention priority and spatial orientation information. Based on the intervention priority, it controls the audible and visual generator to produce audible and visual signals of corresponding intensity and frequency. Based on the yaw and pitch angles, it controls the phase of the phased array antenna array, focusing the audible and visual alarm beams onto the spatial area where the at-risk vehicle is located.

[0083] In this embodiment, intervention priorities are generated based on the threshold amplitude mapping of collision risk entropy, achieving hierarchical alarm processing. The yaw and pitch angles of the alarm beam are calculated based on the projected coordinates of the game equilibrium trajectory, enabling the alarm signal to be directionally projected onto the spatial area where the at-risk vehicle is located. The intervention priority and spatial orientation information are encapsulated into a control signaling frame and sent to the roadside phased array alarm device, driving the phased array alarm device to focus and project the audible and visual alarm beam onto the corresponding spatial area where the at-risk vehicle is located, achieving directional spatial projection of the alarm signal and avoiding interference from omnidirectional alarms to unrelated lanes.

[0084] In one embodiment, reference Figure 6 The system receives initial radar point cloud data from roadside millimeter-wave radar and initial image frame data from a visual sensor. The initial radar point cloud data contains distance, angle, and reflection intensity information for multiple radar points, with each radar point carrying a corresponding radar timestamp. The initial image frame data contains pixel-level image information, with each image frame carrying a corresponding image timestamp. Both radar and image timestamps are generated by a unified clock source from the roadside edge computing unit to ensure time reference consistency.

[0085] Extract the radar timestamp from the initial radar point cloud data and the image timestamp from the initial image frame data. For each radar point cloud frame, extract its corresponding radar timestamp. For each image frame, extract its corresponding image timestamp. The radar timestamps and image timestamps are arranged in chronological order to form a timestamp sequence.

[0086] The time difference between radar timestamps and image timestamps is calculated. Using the data source with the earlier timestamp as the baseline, polynomial fitting interpolation is performed on the moving target trajectory in the data source with the later timestamp. For each pair of adjacent radar timestamps and image timestamps, the time difference between them is calculated. If the radar timestamp is earlier than the image timestamp, the trajectory of the moving target in the image data is interpolated based on the radar data; if the image timestamp is earlier than the radar timestamp, the trajectory of the moving target in the radar data is interpolated based on the image data.

[0087] The polynomial fitting interpolation method is as follows: First, extract the trajectory data of the moving target at multiple consecutive time steps from the reference data source. Then, fit the trajectory data using an nth-degree polynomial to obtain the polynomial expression of the trajectory. Finally, based on the timestamp of the later data source, calculate the trajectory state of the moving target at that timestamp using the polynomial expression. The degree of the polynomial is determined according to the motion characteristics of the moving target, and a quadratic or cubic polynomial is usually used.

[0088] The interpolated moving target trajectory is spatially registered and fused with the data from the reference data source to generate time-synchronized multi-vehicle trajectory time-series data. Spatial coordinate registration refers to transforming the interpolated moving target trajectory from its original coordinate system to the coordinate system of the reference data source. Spatial coordinate registration is achieved through a pre-calibrated coordinate system transformation matrix. The coordinate system transformation matrix includes a rotation matrix and a translation vector, which are determined through calibration experiments.

[0089] The interpolated moving target trajectory is fused with the data from the reference data source in the same coordinate system to generate time-synchronized multi-vehicle trajectory time-series data. During the fusion process, for the same moving target, the trajectory data from the reference data source and the interpolated trajectory data are weighted and averaged to improve the accuracy of the trajectory data. The weighting coefficients are determined based on the reliability of the data source, with higher-reliability data sources assigned larger weights.

[0090] In this embodiment, by calculating the time difference between the radar timestamp and the image timestamp, and using the data source with the earlier timestamp as a benchmark, polynomial fitting interpolation is performed on the moving target trajectory in the data source with the later timestamp, thus eliminating the time asynchrony deviation of the multi-source data. The interpolated moving target trajectory is then spatially registered and fused with the data from the benchmark data source to generate time-synchronized multi-vehicle trajectory time-series data, ensuring the temporal and spatial consistency of the trajectory data input to the prediction model.

[0091] In a preferred embodiment, the roadside multi-source sensing device further includes a lidar system and a vehicle-to-infrastructure (V2I) communication unit. The lidar is used to acquire three-dimensional point cloud data of the vehicles, providing more accurate vehicle contours and position information. The V2I communication unit is used to receive the vehicles' own status information, including their position, speed, acceleration, and steering signals. The multi-vehicle trajectory time-series data fused by the multi-source sensing device also includes information provided by the lidar and V2I communication unit, further improving the accuracy and completeness of the trajectory data.

[0092] In a preferred embodiment, the edge weights of the dynamic spatiotemporal topology graph also consider vehicle type and lane information. Different weight coefficients are set for different types of vehicles to reflect the differences in interaction strength between them. For vehicles in the same lane, the edge weights are increased to reflect stronger interaction relationships within the same lane. For vehicles in adjacent lanes, smaller edge weights are set to reflect weaker interaction relationships between vehicles in adjacent lanes. For vehicles in non-adjacent lanes, the edge weight is zero, meaning there is no edge connection.

[0093] In a preferred embodiment, the spatiotemporal graph convolutional network further includes an attention mechanism. This attention mechanism adaptively adjusts the contribution weights of different neighboring nodes to the features of the central node. Neighboring nodes with stronger interactions are assigned larger attention weights, while those with weaker interactions are assigned smaller attention weights. By introducing this attention mechanism, the spatiotemporal graph convolutional network can more accurately capture important interactions between vehicles, improving the quality of the spatiotemporal interaction features of the vehicle group.

[0094] In a preferred embodiment, the non-cooperative game model also considers differences in driver driving styles. Based on historical vehicle trajectory data, drivers are categorized into three driving styles: conservative, average, and aggressive. Different weighting coefficients are assigned to the payoff function for each style. Conservative drivers have larger weighting coefficients for both collision penalty and longitudinal discomfort costs; aggressive drivers have smaller weighting coefficients for both collision penalty and longitudinal discomfort costs. By considering these differences in driving styles, the process of solving the game equilibrium trajectory more closely reflects the actual driving behavior of different drivers.

[0095] In a preferred embodiment, the calculation of collision risk entropy also considers the influence of vehicle type and relative speed. Different collision risk weighting coefficients are set for different types of vehicles. The collision risk weighting coefficient is larger between large and small vehicles, and smaller between small vehicles. For vehicle pairs with high relative speeds, the collision risk weighting coefficient is increased. By considering the influence of vehicle type and relative speed, the collision risk entropy can more accurately reflect the degree of collision risk under different conditions.

[0096] In a preferred embodiment, the construction of the dynamic adaptive threshold also considers the influence of traffic flow and weather conditions. A traffic flow correction factor is determined based on the current traffic flow of the road segment; the higher the traffic flow, the larger the correction factor. A weather correction factor is determined based on the current weather conditions; the correction factor for rainy, snowy, and foggy days is greater than that for sunny days. Incorporating the traffic flow correction factor and the weather correction factor into the calculation process of the dynamic adaptive threshold allows the threshold to more comprehensively reflect various factors affecting collision risk.

[0097] In a preferred embodiment, the tiered alarm command further includes hazard avoidance information. This hazard avoidance information is determined based on the type and severity of the collision risk, and includes actions such as deceleration, lane keeping, and avoidance. The hazard avoidance information is encapsulated in a control signaling frame and sent to the roadside phased array alarm device. The roadside phased array alarm device conveys the hazard avoidance information to the driver via voice announcement, helping the driver to take timely and appropriate hazard avoidance actions.

[0098] In a preferred embodiment, the roadside phased array alarm device also communicates with the vehicle's onboard terminal. When a graded alarm command is triggered, the roadside edge computing unit sends alarm information and hazard avoidance measure information to the onboard terminal of the at-risk vehicle via the vehicle-to-infrastructure (V2I) communication unit. The onboard terminal displays the alarm information and hazard avoidance measure information to the driver via a display screen and voice system, further improving the effectiveness of the alarm.

[0099] The technical features and corresponding technical effects used in different embodiments are shown in Table 2.

[0100] Table 2 Technical features and effects of different embodiments

[0101] In this embodiment, the technical solution of the present invention is described in detail through multiple embodiments. Each embodiment provides in-depth refinement of specific technical aspects, supplementing specific technical implementation methods and algorithmic logic, and meeting the legal standard of full disclosure. By constructing a dynamic spatiotemporal topology graph and combining it with a spatiotemporal graph convolutional network and a non-cooperative game model, the technical problems of collision prediction lag and alarm failure in multi-vehicle strong interaction scenarios are solved. Through multiple optimization techniques, the accuracy of risk prediction and the reliability of alarms are further improved.

Claims

1. A method for predicting and warning multi-source vehicle collision risks on highways, characterized in that, include: Receive multi-vehicle trajectory time-series data generated by the fusion of multi-source sensing devices on the roadside; A dynamic spatiotemporal topology graph is constructed based on the multi-vehicle trajectory time series data, with vehicles as nodes and vehicle spacing and relative speed as edge weights. Spatiotemporal graph convolutional network is used to extract the spatiotemporal interaction features of vehicle groups in the dynamic spatiotemporal topology graph. The spatiotemporal interaction features of vehicle groups are then input into a non-cooperative game model to calculate the game equilibrium trajectory of each vehicle in the strategy space. Calculate the spatiotemporal conflict degree of the multi-vehicle game trajectory in the prediction time domain, and generate the collision risk entropy based on the time integral of the spatiotemporal conflict degree; When the collision risk entropy exceeds the dynamic adaptive threshold, a graded alarm command carrying intervention priority and spatial orientation information is generated and output to the alarm device.

2. The method for predicting and alarming multi-source vehicle collision risks on highways according to claim 1, characterized in that, The step of constructing a dynamic spatiotemporal topology graph based on the multi-vehicle trajectory time-series data, using vehicles as nodes and vehicle spacing and relative speed as edge weights, includes: The time-series data of the multi-vehicle trajectory is sliced ​​along a time sliding window to obtain the slice data at the current moment; Map each vehicle in the current time slice data to a node feature vector; Spatial affinity is calculated based on the reciprocal of the lateral distance, longitudinal distance and relative speed between vehicles, and the spatial affinity is used as the edge weight to construct an adjacency matrix; When the spatial affinity is lower than the preset affinity threshold, the edge connection between the corresponding nodes is disconnected, and when a new vehicle enters the perception range, the corresponding node and edge connection are added to form a dynamic spatiotemporal topology graph structure that evolves with time steps.

3. The method for predicting and alarming multi-source vehicle collision risks on highways according to claim 1, characterized in that, The step of extracting the spatiotemporal interaction features of vehicle groups in the dynamic spatiotemporal topology graph using a spatiotemporal graph convolutional network includes: In the spatial dimension, graph convolution operations are used to aggregate the node feature vectors in the dynamic spatiotemporal topology graph along the adjacency matrix to obtain the spatial structure interaction features. In the time dimension, a temporal convolution kernel is used to perform convolution operations on the spatial structure interaction features of multiple consecutive time steps along the time axis to capture the temporal evolution pattern of vehicle interaction behavior. The spatial dimension aggregation result and the temporal dimension convolution operation result are combined with feature concatenation and nonlinear mapping to output the spatiotemporal interaction features of the vehicle group that contain the spatiotemporal coupling relationship of the vehicle group.

4. The method for predicting and alarming multi-source vehicle collision risks on highways according to claim 1, characterized in that, The steps of inputting the spatiotemporal interaction characteristics of the vehicle group into a non-cooperative game model to calculate the game equilibrium trajectory of each vehicle in the strategy space include: The revenue function for each vehicle is constructed based on the spatiotemporal interaction characteristics of the vehicle group. The revenue function includes a trajectory smoothing term and a collision penalty term. Candidate actions of each vehicle are sampled in a discrete policy space, and an optimal response policy set is constructed based on the historical policies of other vehicles. By iteratively solving for the Nash equilibrium point of each vehicle under the aforementioned payoff function, the acceleration and angular velocity decision vectors of each vehicle in the game equilibrium state are obtained. Based on the acceleration and angular velocity decision vectors, the current vehicle trajectory state is forward- deduced to generate the game equilibrium trajectory.

5. The method for predicting and alarming multi-source vehicle collision risks on highways according to claim 1, characterized in that, The step of calculating the spatiotemporal conflict degree of the multi-vehicle game trajectory in the prediction time domain, and generating the collision risk entropy based on the time integral of the spatiotemporal conflict degree, includes: Discretize the multi-vehicle game trajectory in the prediction time domain and calculate the bounding box overlap area of ​​any two vehicles at the same time as the spatial conflict component. Calculate the time difference between any two vehicles arriving at the overlapping area of ​​the bounding box as the time conflict component. The product of the spatial conflict component and the temporal conflict component is determined as the spatiotemporal conflict degree; The spatiotemporal conflict degree is integrated along the predicted time domain, and the integration result is mapped to the non-negative real number domain to output the collision risk entropy, which represents the cumulative effect of collision evolution.

6. The method for predicting and alarming multi-source vehicle collision risks on highways according to claim 4, characterized in that, The steps for constructing the revenue function for each vehicle based on the spatiotemporal interaction characteristics of the vehicle group include: Extract the lane departure trend vector and longitudinal acceleration vector from the spatiotemporal interaction features of the vehicle group; A lateral offset cost term is constructed based on the lane departure trend vector, and a longitudinal discomfort cost term is constructed based on the longitudinal acceleration vector; Based on the intersection points of the predicted trajectories of the current vehicle and its neighboring vehicles during the game, a dynamic collision penalty term is constructed. The lateral offset cost, longitudinal ill-adapted cost, and dynamic collision penalty are weighted and summed, with the weight coefficient of the dynamic collision penalty dynamically increasing as the spatial approximation of the predicted trajectory intersection points increases, thus generating the revenue function.

7. The method for predicting and alarming multi-source vehicle collision risks on highways according to claim 5, characterized in that, Before the step of generating a graded alarm instruction when the collision risk entropy exceeds the dynamic adaptive threshold, the step of constructing the dynamic adaptive threshold is also included: Obtain environmental visibility data and road adhesion coefficient for the current road section; A visual perception delay factor is determined based on the environmental visibility data, and a braking distance correction factor is determined based on the road adhesion coefficient. Logarithmic correction calculation is performed on the basic risk threshold based on the visual perception delay factor and braking distance correction factor; The modified base risk threshold is weighted and fused with the historical moving average of the collision risk entropy to output the dynamic adaptive threshold that changes dynamically with the environmental state and historical risk situation.

8. The method for predicting and alarming multi-source vehicle collision risks on highways according to claim 2, characterized in that, The step of mapping each vehicle in the current time slice data to a node feature vector includes: Extract the position coordinates, velocity vectors, and heading angles of each vehicle from the slice data at the current moment; When a target occlusion loss is detected by a multi-source sensing device, a feature transfer matrix is ​​constructed based on the node feature vector of the lost target at a historical moment and the neighbor node features output by the spatiotemporal graph convolutional network. The hidden state of the occluded neighbor node is interpolated and completed using the feature transfer matrix. The interpolated features are then concatenated and dimension-reduced with the position coordinates, velocity vector, and heading angle of the currently unobstructed vehicle to generate the node feature vector after integrity restoration.

9. The method for predicting and alarming multi-source vehicle collision risks on highways according to claim 1, characterized in that, The step of triggering the generation of a graded alarm command carrying intervention priority and spatial orientation information and outputting it to the alarm device includes: The intervention priority is generated based on the over-threshold magnitude mapping of the collision risk entropy; Based on the projection coordinates of the game equilibrium trajectory that generates the collision risk entropy onto the two-dimensional plane, the yaw angle and pitch angle of the alarm beam are calculated as the spatial orientation information. The intervention priority and the spatial orientation information are encapsulated into a control signaling frame and sent to the roadside phased array alarm device. The roadside phased array alarm device analyzes the control signaling frame and focuses and projects the audible and visual alarm beam onto the spatial area where the corresponding risk vehicle is located based on the yaw angle and pitch angle.

10. The method for predicting and alarming multi-source vehicle collision risks on highways according to claim 1, characterized in that, Before the step of receiving the multi-vehicle trajectory time-series data generated by the roadside multi-source sensing devices, the step of time alignment of the multi-source sensing data is also included: Receive initial radar point cloud data output from roadside millimeter-wave radar and initial image frame data output from visual sensor; Extract the radar timestamp from the initial radar point cloud data and the image timestamp from the initial image frame data; Calculate the time difference between radar timestamp and image timestamp, and use the data source with the earlier timestamp as a reference to perform polynomial fitting interpolation on the trajectory of the moving target in the data source with the later timestamp. The interpolated moving target trajectory is spatially registered and fused with the data from the reference data source to generate time-synchronized multi-vehicle trajectory time-series data.

Images