Vehicle collision risk prediction method and device and controller

Abstract

The invention provides a vehicle collision risk prediction method and device and a controller, and relates to the technical field of intelligent driving. The method comprises the following steps: acquiring running state data and road environment data of a vehicle; based on the operation state data and the road environment data, a dynamic interaction network is constructed, and the interaction network takes the vehicle and other traffic participants as nodes and takes the interaction relation between the nodes as edges; extracting the feature representation of each node based on a graph neural network, and further determining the influence score of each node; the influence score of each node is compared with a dynamic threshold value, and sparse attention weight distribution is generated according to a comparison result; and based on the weight distribution, carrying out weighted fusion on the feature representation of each node to obtain a prediction result of the collision risk. According to the invention, the dynamic interaction network is constructed and the hierarchical attention refinement mechanism is introduced, so that the key risk vehicle and the secondary interference target can be accurately distinguished, and the reliability of vehicle collision risk prediction is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent driving, and in particular to a vehicle collision risk prediction method, device and controller. BACKGROUND

[0002] With the rapid development of vehicle intelligence and networking technology, intelligent transportation systems have higher requirements for vehicle active safety capabilities. In complex scenarios such as highway congestion and ramp confluence, the traffic flow density is extremely high, and the dynamic interaction between vehicles is extremely frequent, resulting in a significant increase in collision risk. Such accidents often have serious consequences far beyond a single collision, posing a huge threat to personal safety and the economy. Therefore, the industry's research focus has gradually shifted from collision warning for a single vehicle to accurately and dynamically predicting collision risks in high-density, multi-vehicle interaction scenarios, which has become a core challenge and key technical bottleneck for improving road safety.

[0003] In related technologies, a deep learning model based on an attention mechanism is usually used to predict vehicle collision risks. Specifically, the state data and environmental information of the vehicle and surrounding vehicles are collected through vehicle-mounted sensors and vehicle-to-everything (V2X) communication technology; the information of all interacting vehicles in the scene is weighted and fused using the attention mechanism, aiming to filter out key risk features from the massive data; further, the model outputs a prediction result of the collision risk in the future period of time, which provides a decision basis for subsequent warning or avoidance control. However, there is a problem of low reliability of the vehicle collision risk prediction result. SUMMARY

[0004] The present application provides a vehicle collision risk prediction method, device and controller to improve the problem of low reliability of the vehicle collision risk prediction result in related technologies.

[0005] In a first aspect, the present application provides a vehicle collision risk prediction method, comprising:

[0006] obtaining running state data of the vehicle and road environment data;

[0007] based on the running state data and the road environment data, constructing a dynamic interaction network; wherein the dynamic interaction network takes the vehicle and other road users as nodes and the interaction relationship between the nodes as edges;

[0008] extracting feature representations of the nodes in the dynamic interaction network based on a graph neural network;

[0009] determining influence scores of the nodes based on the feature representations of the nodes, the influence scores being used to quantify the influence degree of the nodes on vehicle collisions;

[0010] The influence score of each node is compared with a dynamic threshold respectively, and a sparse attention weight distribution is generated according to each comparison result, wherein the dynamic threshold is determined based on the running state data and the road environment data;

[0011] Based on the sparse attention weight distribution, the feature representations of the nodes are weighted and fused to obtain a prediction result of the vehicle collision risk.

[0012] In a possible implementation, the feature representation of each node in the dynamic interaction network is extracted based on the graph neural network, including: performing iterative message passing on the dynamic interaction network through a plurality of network layers of the graph neural network; for each node in the nodes, updating the feature representation of the node at the L+1 layer by: aggregating the feature representation of the node at the L layer and the feature representations of the neighbor nodes of the node at the L layer; performing linear transformation on the aggregated result, and processing the linear transformation result by using an activation function to generate the feature representation of the node at the L+1 layer; wherein L is a layer index.

[0013] In a possible implementation, the influence score of each node is determined based on the feature representation of the node, including: multiplying the norm of the feature representation corresponding to the node and the topological centrality metric value of the node in the dynamic interaction network, and taking the multiplication result as the influence score of the node, wherein the norm of the feature representation corresponding to the node is used to quantify the risk urgency of the node causing direct collision, and the topological centrality metric value is used to quantify the key hub degree of the node in the dynamic interaction network.

[0014] In a possible implementation, the running state data includes the vehicle speed of the ego vehicle, and the road environment data includes the traffic flow density, and the dynamic threshold is determined by: inputting the traffic flow density and the vehicle speed of the ego vehicle into a preset linear relationship model to obtain a preliminary threshold; determining a consistency score and a false alarm rate for evaluating the accuracy of the prediction result of the vehicle collision risk based on the prediction result of the vehicle collision risk in a historical setting time period and actual traffic condition data; determining a reward signal based on the consistency score, the false alarm rate and a preset penalty coefficient; and iteratively updating the model parameters of the linear relationship model by using a policy gradient algorithm to maximize the reward signal, to perform online optimization on the preliminary threshold and generate the dynamic threshold.

[0015] In a possible implementation, the vehicle collision risk prediction method further includes: generating a vehicle avoidance control instruction based on the prediction result of the vehicle collision risk, and sending the vehicle avoidance control instruction to an execution mechanism of the vehicle; collecting state feedback data of the vehicle in real time, and determining an execution deviation between the state feedback data and expected state feedback data corresponding to the vehicle avoidance control instruction; if the execution deviation exceeds a preset execution deviation threshold, triggering re-planning of the vehicle avoidance control instruction; meanwhile, taking a sample containing the state feedback data and the execution deviation as incremental data, and using an incremental learning algorithm to update model parameters of the graph neural network online to optimize the accuracy of subsequent vehicle collision risk prediction.

[0016] In a possible implementation, the prediction result of the vehicle collision risk includes a risk probability value of vehicle collision, a predicted risk time interval, a predicted risk spatial coordinate range, and a risk vehicle identification list. The vehicle avoidance control instruction is generated based on the prediction result of the vehicle collision risk, including: determining a target threat vehicle based on the risk vehicle identification list, and determining a minimum safety distance required for the vehicle to achieve complete braking within the predicted risk time interval based on a vehicle speed, the predicted risk time interval, and a maximum deceleration of the vehicle, and taking the minimum safety distance as a safety space boundary. An optimization objective function is constructed with the safety space boundary as a hard constraint. The optimization objective function is configured to minimize the amplitude of the control instruction and optimize a ride comfort index within a future prediction time domain. An optimal control sequence including braking, steering, and power control amounts is obtained by using a model predictive control algorithm to rollingly solve the optimization objective function. The control amount corresponding to the current time in the optimal control sequence is output as the vehicle avoidance control instruction.

[0017] In a possible implementation, in the step of constructing the dynamic interaction network, the weight of each edge in the dynamic interaction network is determined by the following method: for any two nodes in the nodes, the interaction intensity is determined based on the position vector and the velocity vector of the two nodes; the relative acceleration length is determined based on the acceleration vector of the two nodes; the scaling coefficient is determined based on the relative acceleration length; the interaction intensity is multiplied by the scaling coefficient, and the result obtained by the multiplication is taken as the edge weight of the two nodes; wherein the scaling coefficient is positively correlated with the relative acceleration length, and when the relative acceleration length increases, the interaction intensity is amplified.

[0018] In a second aspect, the application provides a vehicle collision risk prediction device, including:

[0019] The acquisition module is configured to acquire running state data of the vehicle and road environment data.

[0020] The dynamic interaction network construction module is configured to construct a dynamic interaction network based on the running state data and the road environment data, wherein the dynamic interaction network takes the ego vehicle and other traffic participants as nodes and takes the interaction relationship between the nodes as edges.

[0021] The feature extraction module is configured to extract feature representations of the nodes in the dynamic interaction network based on a graph neural network.

[0022] The determination module is configured to determine influence scores of the nodes based on the feature representations of the nodes, wherein the influence scores are used to quantify the influence degrees of the nodes on the vehicle collision.

[0023] The comparison module is configured to compare the influence scores of the nodes with dynamic thresholds respectively and generate a sparse attention weight distribution according to the comparison results, wherein the dynamic thresholds are determined based on the running state data and the road environment data.

[0024] The risk prediction module is configured to perform weighted fusion on the feature representations of the nodes based on the sparse attention weight distribution to obtain a prediction result of the vehicle collision risk.

[0025] In a third aspect, the present application provides a controller, comprising a processor and a memory connected with the processor in communication;

[0026] The memory is configured to store computer execution instructions.

[0027] The processor is configured to execute the computer execution instructions stored in the memory to implement the method of any one of the first aspect.

[0028] In a fourth aspect, the present application provides a computer readable storage medium, which stores computer execution instructions, and the computer execution instructions are used to implement the method of any one of the first aspect when executed.

[0029] In a fifth aspect, the present application provides a computer program product, which comprises a computer program, and the computer program is used to implement the method of any one of the first aspect when executed.

[0030] The vehicle collision risk prediction method, device and controller provided in the application obtain running state data and road environment data of a vehicle; a dynamic interaction network is constructed based on the running state data and the road environment data, wherein the dynamic interaction network takes the vehicle and other traffic participants as nodes and the interaction relationship between the nodes as edges; feature representations of the nodes in the dynamic interaction network are extracted based on a graph neural network, and an influence score of each node is determined based on the feature representations of the nodes, the influence score being used to quantify the influence degree of the node on a vehicle collision; the influence scores of the nodes are compared with dynamic thresholds respectively, and a sparse attention weight distribution is generated according to each comparison result, wherein the dynamic thresholds are determined based on the running state data and the road environment data; further, the feature representations of the nodes are weighted and fused based on the sparse attention weight distribution to obtain a prediction result of the vehicle collision risk.

[0031] In this process, by constructing a dynamic interaction network and introducing a hierarchical attention refinement mechanism, the calculation overload and focal point dispersion problem of the traditional attention model in high-density traffic flow are effectively improved. The mechanism extracts deep feature representations of the nodes by using a graph neural network, and then quantitatively evaluates the influence scores of the nodes, and further realizes adaptive attention sparsification screening based on the dynamic thresholds and the influence scores of the nodes, so as to accurately distinguish key risk vehicles from secondary interference targets, while ensuring real-time response, significantly improving the real-time, accuracy and robustness of risk prediction, not only ensuring efficient processing within a limited calculation period, but also providing a highly reliable decision basis for a vehicle avoidance control system by reducing the false positive rate and the false negative rate, thereby effectively ensuring the safety and reliability of the vehicle collision avoidance system in actual complex traffic environments. BRIEF DESCRIPTION OF DRAWINGS

[0032] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the application and, together with the description, serve to explain the principles of the application.

[0033] Figure 1 A flowchart of a vehicle collision risk prediction method provided for an exemplary embodiment of the application;

[0034] Figure 2 Another flowchart of a vehicle collision risk prediction method provided for an exemplary embodiment of the application;

[0035] Figure 3 A structure diagram of a vehicle collision risk prediction device provided for an exemplary embodiment of the application;

[0036] Figure 4 A structure diagram of a controller provided for an exemplary embodiment of the application.

[0037] The present application has been shown and described with reference to the preferred embodiments. Equivalent mechanisms and processes can be used as substitutes for those described and shown, and certain features of the application can be employed without a corresponding use of the others. DETAILED DESCRIPTION

[0038] The exemplary embodiments will be described in detail herein with reference to the attached drawings. The description herein is intended for illustrating the conceptual aspects of the application by way of example and by reference to particular embodiments thereof. It is therefore contemplated that other embodiments, modifications, and variations can be made in the application, as will be suggested by the teachings herein, and will fall within the scope of the application.

[0039] The terms "first", "second", and the like, as used herein do not imply any particular order, but are used for identification only. It is to be understood that the use of the terms "including", "comprising", or "having" and variations thereof herein is intended to encompass the presence of one or more of the stated elements or steps and is not meant to preclude the presence or addition of one or more other elements, steps, features, compositions, and / or data. It is further noted that the claims can be drafted to exclude any elements or steps from the disclosure, regardless of whether or not the elements or steps are specifically disclosed in the claims.

[0040] It is to be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of the relevant data need to comply with relevant laws, regulations and standards, and provide corresponding operation portal for the user to choose authorization or refusal.

[0041] In the related art, when a deep learning model based on an attention mechanism is used to predict vehicle collision risks, the road environment data and environmental information of the vehicle and surrounding vehicles are usually collected by vehicle-mounted sensors and V2X technology, a deep neural network model based on an attention mechanism is used for risk prediction, and a control instruction is generated to drive the vehicle actuators to avoid secondary collisions. The inventors found during research that when applied to ultra-high-density traffic flow, such as highway congestion slow-down or urban expressway ramp merging area, the attention mechanism will have an "attention overload" phenomenon due to the exponential increase in the number of interacting vehicles, highly complex and nonlinear interaction relationships. Specifically, the model's attention resources are crowded out by a large number of concurrent interaction requests, causing its computational load to grow exponentially, and the attention weight distribution module has difficulty accurately distinguishing between "key risk vehicles" that pose a direct threat and "secondary vehicles" that only produce background interference from among dozens or even hundreds of interacting vehicles within a limited real-time computation period, thereby causing a series of serious consequences, including: 1) directly causing the response delay of the risk prediction module, failing to issue a warning within the critical time window before the collision risk occurs, and missing the best opportunity for avoidance operation. 2) may cause distortion of the prediction results, manifested in two extremes: one is the overreaction to secondary vehicles, resulting in a large number of false positives, frequent triggering of unnecessary emergency braking or steering, severely affecting driving comfort and traffic flow efficiency, and even causing new rear-end collision risks; the other is the omission of key risk vehicles, i.e., ignoring the real danger of collision threats, resulting in complete failure of avoidance control. This uncertainty and unreliability of prediction severely restricts the deployment and application of this scheme in real complex scenarios.

[0042] To solve the above problems, the vehicle collision risk prediction scheme provided by the embodiments of the present application introduces a hierarchical attention refinement mechanism, extracts the deep feature representation of each node through a graph neural network, and then quantitatively evaluates the influence score of each node. An adaptive dynamic threshold is used to sparsely filter the global attention, and only the key risk vehicle nodes with an influence score higher than the dynamic threshold are assigned significant attention weights, effectively solving the prediction delay or false positive problem caused by the overload of computational resources and the dispersion of focus of the traditional attention mechanism in high-density traffic flow. Thus, the effect of accurately and real-time identifying core collision risk sources in extremely complex scenarios is achieved, significantly improving the real-time, accuracy and robustness of risk prediction, not only ensuring efficient processing within a limited computation period, but also reducing false positive and false negative rates, providing a highly reliable decision basis for the vehicle avoidance control system, and thus effectively ensuring the safety and reliability of the vehicle collision avoidance system in actual complex traffic environments.

[0043] The technical solutions of the present application and how the technical solutions of the present application solve the above technical problems will be described in detail below with specific examples. The following specific examples can be combined with each other, and the same or similar concepts or processes can not be described again in some examples. The embodiments of the present application will be described below with reference to the drawings.

[0044] Figure 1 A flowchart of a vehicle collision risk prediction method provided for an exemplary embodiment of the present application is shown. As shown in the figure, the vehicle collision risk prediction method comprises the following steps: Figure 1

[0045] S101, obtaining running state data of the vehicle and road environment data.

[0046] For example, the running state data of the vehicle is obtained through the Controller Area Network (CAN) bus of the vehicle, and at the same time, the road environment data is obtained through vehicle-mounted sensors such as millimeter wave radar, laser radar, and V2X communication module. The road environment data includes real-time motion information of other traffic participants (such as vehicles, pedestrians, and non-motor vehicles, etc.) around the vehicle, as well as static environment information. These data together constitute the original input of the system perception layer.

[0047] S102, constructing a dynamic interaction network based on the running state data and the road environment data; wherein the dynamic interaction network takes the vehicle and other traffic participants as nodes, and the interaction relationship between each node as an edge.

[0048] The dynamic interaction network is a traffic scene mathematical model constructed in the form of a dynamic graph, which is used to abstract traffic participants and their interaction relationships into a topological structure of nodes and edges. For example, the vehicle and each perceived traffic participant (such as a vehicle in front, a vehicle cutting in from the side, etc.) are modeled as a node in the dynamic interaction network; the distance, relative speed, etc. between any two nodes are calculated, and if they meet the preset interaction condition (such as a distance less than 100 meters), an edge is established between the two nodes. The weight (i.e. the strength of the interaction relationship) of the edge can be assigned by an initial interaction strength function, for example, the closer the distance, the greater the speed difference, the higher the weight of the edge. The dynamic interaction network is dynamic, and the node attributes and edge relationships are dynamically updated at each calculation period as the vehicle moves.

[0049] S103, extracting feature representation of each node in the dynamic interaction network based on a graph neural network.

[0050] ​The graph neural network is a deep learning model specially used for processing graph structure data; and the feature representation refers to a high-dimensional feature vector of each node containing its own state and neighbor node information after being processed by the graph neural network. Accordingly, the constructed dynamic interaction network is input into a pre-trained graph neural network, which works through an iterative message passing mechanism. Specifically, each node (vehicle) receives the feature information of its neighbor nodes (vehicles connected by edges), and aggregates the information with its own features to update its own feature representation. After multiple iterations, each node obtains a deep feature vector reflecting its context information in the entire interaction network.

[0051] In S104, an influence score of each node is determined based on the feature representation of the node, and the influence score is used to quantify the influence degree of the node on the vehicle collision.

[0052] The influence score is a scalar value used to comprehensively quantify the contribution degree or threat level of a certain vehicle node to the entire collision risk scenario. Accordingly, for each node in the dynamic interaction network, the influence score of the node is determined based on the feature representation of the node obtained in S103.

[0053] In some embodiments, determining the influence score of each node based on the feature representation of the node includes taking the norm of the feature representation of the node as the influence score of the node.

[0054] For example, the influence score of the i-th node is satisfies the following formula:

[0055]

[0056] wherein, is the feature representation of the node; is the norm of the feature representation of the node, used to quantify the risk urgency of the node causing direct collision.

[0057] In some embodiments, determining the influence score of each node based on the feature representation of the node includes multiplying the norm of the feature representation of the node by a topological centrality measure value of the node in the dynamic interaction network, and taking the result of the multiplication as the influence score of the node, wherein the topological centrality measure value is used to quantify the key hub degree of the node in the dynamic interaction network for spreading collision.

[0058] For example, the influence score of the i-th node is satisfies the following formula:

[0059]

[0060] wherein, is a topological centrality measure value of the node in the dynamic interaction network, used to quantify the degree of the node as a key pivot in the propagation of the collision in the dynamic interaction network.

[0061] It should be noted that the above topological centrality measure value is determined according to the topological connection structure of the real-time dynamic interaction network constructed in the above step S102, and is a quantitative description of the degree of the i-th node as a key in the corresponding dynamic graph of the dynamic interaction network. The specific value can be obtained by calculating the degree centrality, betweenness centrality or eigenvector centrality of the node in the dynamic graph, so as to identify the key vehicle node that may play a pivotal or bridge role in the propagation of the collision risk from the network structure level.

[0062] S105, compare the influence scores of each node with the dynamic threshold value respectively, and generate a sparse attention weight distribution according to each comparison result, wherein the dynamic threshold value is determined based on the running state data and the road environment data.

[0063] wherein, the dynamic threshold value is a numerical threshold dynamically adjusted according to the real-time traffic scene change; the sparse attention weight distribution means that only a few key nodes are assigned significant weights by screening, and the weights of most nodes are suppressed to a distribution state close to zero.

[0064] For example, compare the influence scores of all nodes obtained in S104 with the dynamic threshold value For example, according to the vehicle density of the current road and the speed of the vehicle, the dynamic threshold value is determined in real time; only the nodes with influence scores exceeding this threshold value (i.e. truly high-risk key vehicles) will be assigned higher attention weights, and the weights of the remaining nodes will be significantly suppressed. This process is called "sparsification", which is used to force the system to concentrate limited computing resources on a small number of most dangerous targets.

[0065] S106, based on the sparse attention weight distribution, the feature representations of each node are weighted and fused to obtain the prediction result of the vehicle collision risk.

[0066] For example, based on the sparse attention weight distribution, the feature representations of each node are weighted and summed, and the weighted scene features are input into a prediction layer (such as a fully connected neural network), and the risk probability value is output through a Softmax function. The risk probability value satisfies the following formula, for example:

[0067]

[0068] wherein, is an output layer weight matrix, sparse attention weights of the node, represent a risk probability value of collision. Further, a key risk vehicle set is identified through a feature alignment mechanism to form a complete risk prediction result, thereby providing a basis for subsequent adaptive avoidance control.

[0069] The vehicle collision risk prediction method provided by the embodiments of the present application comprises the following steps: obtaining running state data and road environment data of a vehicle; constructing a dynamic interaction network based on the running state data and the road environment data, wherein the dynamic interaction network takes the vehicle and other traffic participants as nodes and the interaction relationship between the nodes as edges; extracting feature representations of the nodes in the dynamic interaction network based on a graph neural network, and determining influence scores of the nodes based on the feature representations of the nodes, wherein the influence scores are used to quantify the influence degree of the nodes on vehicle collision; comparing the influence scores of the nodes with dynamic thresholds respectively, and generating a sparse attention weight distribution according to the comparison results, wherein the dynamic thresholds are determined based on the running state data and the road environment data; and further weighting and fusing the feature representations of the nodes based on the sparse attention weight distribution to obtain a prediction result of vehicle collision risk. In this process, by constructing a dynamic interaction network and introducing a hierarchical attention refinement mechanism, the calculation overload and focal point dispersion problems of traditional attention models in high-density traffic flow are effectively improved. This mechanism extracts deep feature representations of the nodes by using a graph neural network, and then quantitatively evaluates the influence scores of the nodes. Further, adaptive attention sparsification screening is realized based on the dynamic thresholds and the influence scores of the nodes, so that key risk vehicles and secondary interference targets can be accurately distinguished. While ensuring real-time response, the real-time accuracy and robustness of risk prediction are significantly improved, not only ensuring efficient processing within a limited calculation period, but also providing a highly reliable decision basis for vehicle avoidance control systems by reducing false positive rates and false negative rates, thereby effectively ensuring the safety and reliability of the vehicle collision avoidance system in actual complex traffic environments.

[0070] In some embodiments, the running state data comprises a vehicle speed of the vehicle, the road environment data comprises a traffic flow density, and the dynamic threshold is determined by the following method: inputting the traffic flow density and the vehicle speed of the vehicle into a preset linear relationship model to obtain a preliminary threshold; determining a consistency score and a false positive rate for evaluating the accuracy of the prediction result of the vehicle collision risk based on the prediction result of the vehicle collision risk and actual traffic condition data within a historical setting time period; determining a reward signal based on the consistency score, the false positive rate and a preset penalty coefficient; and iteratively updating model parameters of the linear relationship model by using a policy gradient algorithm to maximize the reward signal, so as to perform online optimization on the preliminary threshold and generate the dynamic threshold.

[0071] For example, the running state data includes the vehicle speed, acceleration, steering angle, yaw rate, etc. of the ego vehicle; and the road environment data includes the position, speed, acceleration, heading angle of surrounding vehicles, and the road curvature, road adhesion coefficient, and traffic flow density, etc.

[0072] Optionally, after obtaining the running state data of the ego vehicle and the road environment data, the method further comprises preprocessing the data set containing the running state data and the road environment data performing a preprocessing operation, the preprocessing operation including but not limited to data cleaning, outlier removal, timestamp alignment, and coordinate system unification, etc., to obtain a standardized and integrated data set through the preprocessing operation .

[0073] For example, first, the data set perform data cleaning and outlier removal, which can specifically identify abnormal points that do not conform to physical constraints and remove them through statistical detection, and at the same time, filter the sensor noise and smooth it, for example, using a weighted moving average, which satisfies the following formula:

[0074]

[0075] wherein, is the smoothed signal, is the original observation value, is the weight coefficient, and m is the sliding window radius, which ensures the continuity of the signal and reduces the interference of high-frequency noise on subsequent modeling.

[0076] Correspondingly, after data cleaning, timestamp alignment is performed to synchronize all cleaned data to a unified time reference, for example, by resampling data streams of different frequencies through interpolation, so that at time t, the vehicle state, environmental perception, and V2X communication information all correspond to consistent data snapshots. After timestamp alignment, coordinate system unification is performed to unify the position data from different sources to the ego vehicle coordinate system, and the conversion formula is, for example:

[0077]

[0078] wherein, is the position vector converted to the ego vehicle coordinate system, is the original global coordinate, is the position of the ego vehicle in the global coordinate, is the rotation matrix defined by the heading angle of the ego vehicle . This conversion mechanism ensures that the position information of all road users can be uniformly and consistently expressed in the ego vehicle coordinate system, providing an accurate spatial reference for subsequent interaction relationship modeling.

[0079] Further, data fusion is performed to integrate the observations of different sensors and V2X. By adopting a combination of weighted average and Kalman filtering, the processing accuracy of multi-source heterogeneous data can be effectively improved. The weighted average is used for preliminary fusion of redundant information of the same type, and the Kalman filtering realizes recursive estimation by using the system state transition model and the observation model. The corresponding formulas are as follows:

[0080]

[0081] wherein, is the state estimation value, is the prior estimation, is the observation vector, and H is the observation matrix; is the Kalman gain, which is determined by the covariance ratio of the system noise and the observation noise. Through this process, the fusion results are unified, formatted and standardized, forming a data set with consistency and integrity , thereby providing a stable and reliable input data basis for subsequent risk prediction.

[0082] For example, the traffic flow density and the vehicle speed extracted from the data set are taken as inputs, and the dynamic threshold is output, which is used to filter the impact score in the hierarchical sparse attention calculation. The processing process first performs preliminary calculation based on a linear model, satisfying the following formula:

[0083]

[0084] wherein, is the traffic density weight parameter, which reflects the sensitivity of the traffic flow density to the threshold; is the vehicle speed weight parameter, which reflects the adjustment of the vehicle speed to the threshold; b is the bias term, which provides baseline correction when the sparsity is too low or too high; is the ratio of the number of vehicles in a unit road section to the road capacity, which describes the congestion degree of the traffic scene, i.e., the traffic flow density; is the longitudinal speed of the vehicle, which describes the motion state of the vehicle in the current traffic environment. This linear relationship ensures that the threshold is increased in the scene with high traffic flow density or high vehicle speed, so that the attention calculation tends to be sparse, thereby focusing on the most critical risk vehicles and relieving the feature redundancy problem caused by the increase in scene complexity.

[0085] ​Furthermore, a reinforcement learning fine-tuner is introduced to achieve dynamic online adjustment of the threshold. The fine-tuner generates a reward signal based on the prediction accuracy over a recent period, such as the last 500 seconds. , Satisfy the following formula:

[0086]

[0087] Among them, the consistency score reflects the degree of consistency between the risk prediction results and the actual collision trend; the false alarm rate is the proportion of false high-risk alarms generated in the prediction. This is the penalty coefficient, used to balance the trade-off between accuracy and robustness in the model. The reinforcement learning fine-tuner updates parameters using the policy gradient algorithm. The adjustment range makes During operation, it can continuously adjust based on the environment and prediction performance. Among these adjustments, the policy gradient algorithm is a core method in reinforcement learning that directly optimizes the policy itself. It is defined as an algorithm that iteratively improves policy quality by calculating the gradient of the policy performance index (expected total reward) with respect to policy parameters and updating the policy parameters along the gradient direction. This mechanism, through iterative learning of the reward signal, gradually forms an adaptive threshold adjustment policy for different traffic densities and vehicle speeds, making attention allocation more efficient. It can suppress invalid interference nodes in high-density, high-speed scenarios and retain more potential risk information in low-density, low-speed scenarios, thereby improving prediction stability and reducing latency, further enhancing the real-time performance and reliability of risk prediction in complex and dynamic traffic environments.

[0088] In some embodiments, feature representations of each node in a dynamic interaction network are extracted based on a graph neural network, including: iterative message passing through multiple network layers of the graph neural network; for each node, the feature representation of the node at layer L+1 is updated in the following way: aggregating the node's own feature representation at layer L and the feature representations of each of the node's neighboring nodes at layer L; performing a linear transformation on the aggregated result and processing the linear transformation result with an activation function to generate the feature representation of the node at layer L+1; where L is the layer index.

[0089] For example, the first step is to execute the dynamic interactive network construction sub-step, specifically, at time t, the dynamic interactive network construction sub-step is executed. Represented as a dynamic image ,in, For each set of nodes, the feature vector of each node is... By speed acceleration and position vector Composition, edge set edge weight The interaction intensity function is determined by the following formula:

[0090]

[0091] wherein, and denote the position vector of node i and node j, and denote the velocity vector of the two nodes, is a constant to prevent the denominator from being zero; the function describes the interaction intensity between nodes through the position distance and the velocity difference, so that nodes with close distance and large velocity difference, such as vehicles, contribute more significantly to potential risks.

[0092] Correspondingly, the graph neural network feature extraction sub-step is performed to input the dynamic graph into the graph neural network, and the node features are updated through iterative message passing, which satisfies the following formula:

[0093]

[0094] wherein, is the neighbor set of node i, and are trainable weight matrices, is an activation function, is the attention weight corresponding to the Lth layer, is the feature representation of the node in the L+1th layer. This step learns the high-dimensional representation of the node in the complex interaction scenario by aggregating the neighbor features and the self features, thereby preserving the dynamic dependence relationship between vehicles.

[0095] The embodiments of the present application effectively improve the representation ability of vehicle interaction features through the multi-layer iterative message passing mechanism of the graph neural network. This structure aggregates the node's own features and neighbor node features, so that each node can fuse local topological relationship information; after linear transformation and activation function processing, a high-dimensional feature representation containing multi-level interaction context is gradually constructed. This hierarchical feature learning mechanism enables the model to capture complex spatio-temporal dependence relationships in traffic scenes, lays a solid feature foundation for subsequent accurate evaluation of node influence scores, and thereby significantly enhances the context awareness ability of risk prediction.

[0096] In some embodiments, the vehicle collision risk prediction method further comprises: generating a vehicle avoidance control instruction based on the prediction result of the vehicle collision risk, and sending the vehicle avoidance control instruction to the execution mechanism of the vehicle; collecting state feedback data of the vehicle in real time, and determining an execution deviation between the state feedback data and the expected state feedback data corresponding to the vehicle avoidance control instruction; if the execution deviation exceeds a preset execution deviation threshold, triggering re-planning of the vehicle avoidance control instruction; at the same time, taking the sample containing the state feedback data and the execution deviation as incremental data, and using an incremental learning algorithm to update the model parameters of the graph neural network online to optimize the accuracy of subsequent vehicle collision risk prediction.

[0097] For example, when the prediction result of the vehicle collision risk contains a "high risk" prediction result, a specific vehicle avoidance control instruction, such as "apply 40% braking force", is generated based on the "high risk" prediction result. The instruction is sent to the execution mechanism of the vehicle, such as the Electronic Stability Program (ESP) or the Electric Power Steering (EPS), through the vehicle CAN bus in real time. At the same time of issuing the instruction, the state feedback data of the vehicle, such as the actual deceleration of the vehicle, is collected in real time through the vehicle sensors, such as the wheel speed sensor or the Inertial Measurement Unit (IMU). Further, the actual deceleration is compared with the expected deceleration (calculated from the "40% braking force" instruction) to calculate the execution deviation (for example, the actual deceleration is only 80% of the expected deceleration).

[0098] Correspondingly, based on the preset execution deviation threshold, such as 15% of the expected deceleration, if the execution deviation exceeds the execution deviation threshold, it indicates that the current control instruction fails to achieve the expected effect (possibly due to wet road surface, insufficient tire grip, etc.), and the control system will immediately trigger re-planning, for example, taking the current sensor feedback and deviation information as new input, quickly solving a new and more appropriate vehicle avoidance control instruction (for example, increasing the braking force to 55%), to replace the original vehicle avoidance control instruction, to ensure immediate safety.

[0099] In addition, while monitoring the control execution, the state feedback data (actual deceleration) and the execution deviation and other information are bound with the risk prediction scene data (such as the dynamic graph corresponding to the dynamic interaction network at that time) that leads to this control to form an incremental data sample; further, using an incremental learning algorithm such as an online gradient descent algorithm, the model parameters of the graph neural network are fine-tuned using the sample to update the model parameters of the graph neural network online, thereby optimizing the accuracy of subsequent vehicle collision risk prediction.

[0100] This application's embodiments effectively improve the system's real-time adaptability and long-term reliability by establishing a closed-loop feedback mechanism from decision-making to execution. By comparing the expected effect of the vehicle evasion control command with the actual vehicle state's execution deviation, command replanning can be triggered immediately upon detecting a risk of control failure, ensuring the timely and effective execution of evasion actions. Simultaneously, the execution deviation data is used as incremental learning samples to optimize the graph neural network online, ensuring continuous learning and self-correction from actual interactions. This dual optimization mechanism not only guarantees the accuracy of individual control responses but also improves prediction accuracy and generalization ability in complex traffic scenarios through continuous accumulation of practical experience.

[0101] In some embodiments, the predicted vehicle collision risk includes a vehicle collision probability value, a predicted risk time interval, a predicted risk spatial coordinate range, and a list of risk vehicle identifiers. Based on the predicted vehicle collision risk, a vehicle evasion control command is generated, including: identifying target threat vehicles based on the risk vehicle identifier list, and determining the minimum safe distance required for the vehicle to achieve full braking within the predicted risk time interval based on the vehicle speed, the predicted risk time interval, and the vehicle's maximum deceleration, and using the minimum safe distance as the safe space boundary; constructing an optimization objective function with the safe space boundary as a hard constraint, the optimization objective function being configured to minimize the amplitude of the control command and optimize the ride comfort index within a future predicted time domain; solving the optimization objective function using a model predictive control algorithm to obtain an optimal control sequence including braking, steering, and power control quantities; and outputting the control quantity corresponding to the current moment in the optimal control sequence as the vehicle evasion control command.

[0102] For example, the prediction results of vehicle collision risk For input, the result contains risk probability values. Risk time window (i.e., the predicted risk time interval) Potential collision zone (i.e., the predicted risk spatial coordinate range) And key risk vehicle identifiers (i.e., a list of risk vehicle identifiers). Output the optimal control command sequence Among them, real-time control quantity It was directly issued to the vehicle execution agency. Specifically, through analysis... and To identify the most threatening vehicles and their duration of operation, and to calculate the safety space boundary in conjunction with the vehicle's dynamic constraints. The safety boundary is defined as a minimum safety distance function, which specifically satisfies the following formula:

[0103]

[0104] wherein, is the current vehicle speed, is the maximum deceleration that the vehicle can achieve under the current road adhesion condition, is the risk time window, is the minimum longitudinal distance required to avoid a collision at a given speed and braking capability, which characterizes the physical safety range that the vehicle can achieve within a limited time.

[0105] Accordingly, on the basis of the above risk situation analysis, a prediction control strategy generation sub-step is entered to An optimization objective function is constructed as a constraint condition, which satisfies the following formula:

[0106]

[0107] wherein, represents the control input vector at the kth step of the prediction time domain, including the control amount of braking, steering, and power system; The term is used to limit the amplitude of the control instruction to prevent over-activation. is a comfort index function used to describe the influence of acceleration change rate and lateral acceleration on passenger comfort; w is a comfort weight coefficient used to balance safety and comfort; M is the length of the prediction time domain used to determine the time range of the controller's foresight. By solving the optimization objective function with constraints, the optimal control sequence , satisfies the following formula:

[0108]

[0109] wherein, is the optimal control input at the current time. Further, a command execution and feedback sub-step is performed, i.e., the immediate control amount is sent to the braking, steering, and power systems through the vehicle-mounted CAN bus to ensure that the vehicle responds quickly to risks under physical constraints and safety boundary conditions; at the same time, the running state of the vehicle is collected in real time to update the data set in step S101 , so that the vehicle collision risk prediction and vehicle avoidance control form a closed loop. This mechanism ensures that the system can adjust the control strategy according to the risk situation in a dynamic traffic scenario, effectively avoiding secondary collisions, and maintaining an optimal balance between safety and comfort, thereby improving overall driving safety and robustness.

[0110] In some embodiments, in the step of constructing the dynamic interaction network, the weight of each edge in the dynamic interaction network is determined by: for any two nodes in the nodes, determining an interaction intensity based on the position vector and the velocity vector of the two nodes; determining a relative acceleration modulus based on the acceleration vector of the two nodes; determining a scaling coefficient based on the relative acceleration modulus; multiplying the interaction intensity and the scaling coefficient, and taking the result of the multiplication as the edge weight of the two nodes; wherein the scaling coefficient is positively correlated with the relative acceleration modulus, and when the relative acceleration modulus increases, the interaction intensity is amplified.

[0111] For example, for any two nodes, such as node i and node j, in the nodes, the feature representation of node i and node j is respectively denoted as and , and the input is , which contains elements such as velocity, acceleration, and position, and the output is the enhanced interaction intensity , which is used to update the edge weight of the dynamic graph. The processing process is realized by extending the original interaction intensity function, and the enhanced interaction intensity , that is satisfies the following formula:

[0112]

[0113] wherein is the original interaction intensity function, which satisfies the following formula:

[0114]

[0115] wherein and denote the position vector of node i and node j, and denote the velocity vector of the two nodes, is a constant to prevent the denominator from being zero, and the function describes the interaction intensity between nodes through the position distance and the velocity difference, so that nodes with close distance and large velocity difference, such as vehicles, contribute more significantly to potential risks; the extended function introduces an acceleration difference factor on this basis, wherein and denote the acceleration vector of node i and node j, respectively, and are extracted from the data set . denotes the relative acceleration modulus, which is used to quantify the difference in acceleration or deceleration between the two vehicles; is a coefficient for adjusting the weight of the acceleration difference, and the larger the value is, the more significant the amplification effect of the acceleration difference on the interaction intensity.

[0116] The enhanced computing mechanism considers the potential risk brought by the acceleration difference in addition to the vehicle position and speed characteristics. When the acceleration change directions of two vehicles are opposite or the difference is large, the interaction intensity significantly increases, thereby highlighting the impact of emergency dynamic changes on the collision risk and prompting the edge weight of the dynamic interaction network to more sensitively reflect the risk propagation characteristics of the sudden event in the traffic flow. This enables the subsequent graph neural network to pay more attention to high-risk interaction relationships when performing feature aggregation, further improving the response speed and accuracy of vehicle collision risk prediction in complex traffic environments.

[0117] It should be noted that the above series of steps from constructing a dynamic interaction network to outputting a risk prediction result can also be a closely connected and indivisible organic whole, which together constitutes the complete execution process of the vehicle collision risk prediction model. The model forms an end-to-end computing framework from multi-source data to risk prediction through the synergistic effect of dynamic graph construction, graph neural network feature extraction, influence score calculation, dynamic threshold screening, and sparse attention weighted fusion. This integrated architecture not only ensures the data consistency between the algorithm modules, but more importantly, it realizes accurate perception and efficient calculation of high-density traffic scenarios through a hierarchical attention refinement mechanism, ultimately enabling the model to meet the requirements of real-time, accuracy, and robustness.

[0118] Correspondingly, the training process of the vehicle collision risk prediction model starts with an input data set which is derived from historical traffic collision scenarios and multi-dimensional data streams generated by virtual simulation, covering vehicle state vector , environmental perception vector , and V2X interaction features . Through the aforementioned preprocessing operations, the data set is processed into a standardized feature set with a unified time reference and coordinate system. The vehicle collision risk prediction model is composed of an input layer, multiple hidden layers, and an output layer. The hidden layers include graph neural network layers to capture the topological relationship between vehicles and the environment. The nodes represent vehicle state features, and the edges represent vehicle interactions. The node representation is updated through a message passing mechanism, which satisfies the following formula:

[0119]

[0120] wherein, is the feature representation of node i in the Lth layer, is the feature representation of node i in the (L+1)th layer, denotes the set of neighbor nodes connected to node i, is the weight matrix of the Lth layer, is the bias term, is a nonlinear activation function. Accordingly, after the feature extraction of the graph neural network, the node features are input into the attention mechanism layer, which calculates the attention weight between nodes through a learnable scoring function, and the attention weight satisfies the following formula:

[0121]

[0122] wherein, is the attention weight between node i and neighbor node j, a is a learnable parameter vector, and || represents the vector concatenation operation; the output layer is mapped to two tasks, which are collision occurrence probability prediction and future risk level numerical regression, respectively; the loss function is a weighted combination of cross-entropy loss for classification task and mean square error for regression task:

[0123]

[0124] wherein, , represents the true collision label of sample i, is the probability predicted by the model, is the true risk level value, is the predicted value, and N is the total number of training samples, and are balance coefficients. The optimizer uses the Adam algorithm to iteratively update the parameters according to the gradient information, and the update rule satisfies the following formula:

[0125]

[0126] wherein, is the model parameter of the t-th iteration, is the learning rate, and are the bias correction values of the first and second moment estimates, respectively, is a constant to prevent the denominator from being zero. During the training process, an early stopping strategy is introduced, which specifically monitors the change of the validation set loss, and terminates the training when no improvement is made for a certain number of consecutive rounds, to avoid overfitting. Finally, the converged weight matrix and bias parameter set are obtained, ensuring good generalization ability in diverse traffic environments and providing reliable risk prediction input for subsequent adaptive avoidance control.

[0127] Correspondingly, the execution and monitoring process of the vehicle avoidance control instruction takes the optimal control command sequence as input, which includes the brake control amount , the steering control amount , and the throttle control amount , which are converted into The low-level instructions conforming to the actuator interface are converted and issued in real time to the braking system, steering actuator and powertrain via the CAN bus, ensuring that the instruction transmission delay and jitter are minimized. Further, the sensor system monitors the execution process in real time to obtain a vehicle state feedback vector , and compares it with the expected command at each moment. The execution deviation satisfies the following formula:

[0128]

[0129] wherein, is the execution deviation at time t, is the actual output control quantity of the actuator, is the target control quantity at the moment, and the norm of the vector difference reflects the control accuracy. If exceeds the execution deviation threshold , a dynamic adjustment mechanism is immediately triggered, that is, the current sensor feedback and deviation information are re-fed into the relevant calculation module to obtain new vehicle evasion control instructions to replace the original vehicle evasion control instructions, thereby maintaining the execution stability.

[0130] Correspondingly, all the actual execution effect feedback obtained through monitoring will be returned to the data acquisition and fusion link as real-time closed-loop feedback, continuously correcting the accuracy of the input data and improving the reliability of subsequent prediction and control, thereby forming a closed-loop link of perception, prediction and execution, so that the entire system maintains high precision and high robustness when facing complex traffic environments.

[0131] By way of example, the system closed-loop feedback mechanism takes the execution effect feedback and the acquired real-time data as inputs, and outputs the model parameters of the updated vehicle collision risk prediction model and the optimized control strategy parameters, including the Model Predictive Control (MPC) objective function weight w. The processing process first performs performance analysis on the feedback data, specifically by calculating the execution deviation and the vehicle state deviation index to identify the control accuracy decrease, prediction error increase or abnormal behavior node, thereby locating the system performance bottleneck. Subsequently, an incremental learning algorithm is applied to update the vehicle collision risk prediction model online, and the newly acquired real-time data and the feedback information are fused to adjust the weight matrix and the bias term of the graph neural network and the attention layer, and the update formula is as follows:

[0132]

[0133] in, For the current model parameter set, For learning rate, The incremental learning loss function combines cross-entropy and mean squared error to measure the deviation of new data from the model's predictions, ensuring that the model can quickly adjust after receiving new samples and adapt to the dynamic traffic environment.

[0134] Simultaneously, the control strategy parameters are optimized based on the feedback results, specifically the comfort weights in the MPC objective function J. Perform adaptive adjustments:

[0135]

[0136] Among them, by implementing deviations Passenger comfort index In combination, an incremental learning algorithm is used to solve for the new weights w', so that the control sequence is within the safety constraints. The closed-loop mechanism not only utilizes execution feedback to achieve real-time control correction, but also drives the online evolution of vehicle collision risk prediction models and control strategies, endowing the system with continuous self-learning and improvement capabilities. This ensures high performance levels in prediction accuracy, response speed, and robustness in long-term, complex, and ever-changing traffic environments.

[0137] In addition, the performance monitoring and optimization process takes system operation logs and real-time performance metrics as input, including risk prediction accuracy. With control response time The output is the optimized system configuration parameters, including dynamic thresholds. coefficient With bias The processing begins by collecting and analyzing data from the runtime logs to obtain key performance indicators. The average prediction error satisfies the following formula:

[0138]

[0139] in, Let i be the true risk probability of sample i. The model predicts the value, where N is the total number of samples, and the control response delay is also calculated. This refers to the time required for a command to be issued to the vehicle actuator and for it to reach the desired state, used to measure the system's real-time performance. These performance metrics are then input into a multi-objective optimization algorithm, and the dynamic threshold formula is adjusted accordingly. In The b parameter, along with other parameters, allows for a reduction in response latency while maintaining high prediction accuracy, thus balancing the real-time performance and precision of prediction and control.

[0140]

[0141] This optimization algorithm determines the optimal parameter combination by iteratively searching and weighing different indicators in a weighted objective function. ), and combine the optimal parameters. Update the system configuration. Furthermore, continuously monitor hardware performance metrics and software health, and automatically trigger maintenance operations (such as sensor recalibration and cache data cleanup) when performance deviations are detected. Automated tools complete closed-loop optimization from data acquisition and performance metric calculation to model parameters, control strategies, and dynamic thresholds, forming a complete self-maintenance system. This automated operation and maintenance mechanism significantly reduces the need for manual intervention while ensuring long-term system stability, fundamentally improving the overall system performance and operational reliability.

[0142] Figure 2 Another flowchart illustrating the vehicle collision risk prediction method provided for an exemplary embodiment of this application is shown. Figure 2 As shown, the vehicle collision risk prediction method includes the following steps:

[0143] S201. Obtain the vehicle's operating status data and road environment data.

[0144] For example, operational status data includes vehicle speed, acceleration, steering angle, and yaw rate; road environment data includes the position, speed, acceleration, and heading angle of other traffic participants, as well as road curvature, road surface adhesion coefficient, and traffic flow density.

[0145] Optionally, after acquiring the vehicle's operating status data and road environment data, the process also includes acquiring a dataset containing the operating status data and road environment data. Preprocessing operations are performed, including but not limited to data cleaning, outlier removal, timestamp alignment, and coordinate system unification, to obtain a standardized and integrated dataset. .

[0146] S202. Based on operational status data and road environment data, construct a dynamic interactive network.

[0147] Among them, the dynamic interactive network is a mathematical model of traffic scenarios constructed in the form of a dynamic graph, with the vehicle itself and other traffic participants as nodes, and the interaction relationships between the nodes as edges.

[0148] S203. Extract feature representations of each node in a dynamic interactive network based on graph neural networks.

[0149] In some embodiments, the feature representation of each node in the dynamic interaction network is extracted based on a graph neural network, including: performing iterative message passing on the dynamic interaction network through multiple network layers of the graph neural network; for each node in the nodes, updating the feature representation of the node at the L+1 layer by: aggregating the feature representation of the node at the L layer and the feature representation of each neighbor node of the node at the L layer; performing linear transformation on the aggregated result, and processing the linear transformation result by using an activation function to generate the feature representation of the node at the L+1 layer; wherein L is a layer index.

[0150] In S204, the influence scores of the nodes are determined based on the feature representations of the nodes, and the influence scores are used to quantify the influence degree of the nodes on the vehicle collision.

[0151] Specifically, the norm of the feature representation corresponding to the node is multiplied by the topological centrality metric value of the node in the dynamic interaction network, and the result of the multiplication is taken as the influence score of the node, wherein the norm of the feature representation corresponding to the node is used to quantify the risk urgency of the node causing direct collision, and the topological centrality metric value is used to quantify the key hub degree of the node in the dynamic interaction network for spreading collision.

[0152] In S205, the influence scores of the nodes are compared with a dynamic threshold respectively, and a sparse attention weight distribution is generated according to each comparison result, wherein the dynamic threshold is determined based on the running state data and the road environment data.

[0153] In S206, the feature representations of the nodes are weighted and fused based on the sparse attention weight distribution to obtain a prediction result of the vehicle collision risk.

[0154] For example, the feature representations of the nodes are weighted and summed based on the sparse attention weight distribution, and the weighted scene features are input into a prediction layer (such as a fully connected neural network) to output a risk probability value through a Softmax function, wherein the risk probability value satisfies the following formula, for example:

[0155]

[0156] wherein, is an output layer weight matrix, is the sparse attention weight of the node, represents the risk probability value of collision. Further, a key risk vehicle set is identified through a feature alignment mechanism to form a complete risk prediction result, thereby providing a basis for subsequent adaptive avoidance control.

[0157] In S207, a vehicle avoidance control instruction is generated based on the prediction result of the vehicle collision risk, and the vehicle avoidance control instruction is sent to an execution mechanism of the vehicle.

[0158] Specifically, a target threat vehicle is determined based on a risk vehicle identification list, and a minimum safety distance required for the ego vehicle to achieve full braking within a predicted risk time interval is determined based on a speed of the ego vehicle, the predicted risk time interval, and a maximum deceleration of the ego vehicle, and the minimum safety distance is taken as a safety space boundary; an optimization objective function is constructed with the safety space boundary as a hard constraint, the optimization objective function being configured to minimize a magnitude of a control instruction and optimize a ride comfort index within a future predicted time domain; an optimal control sequence including braking, steering, and power control amounts is obtained by rolling solving the optimization objective function through a model predictive control algorithm; and a control amount corresponding to a current time in the optimal control sequence is taken as a vehicle evasion control instruction output.

[0159] S208, state feedback data of the ego vehicle is collected in real time, and an execution deviation between the state feedback data and expected state feedback data corresponding to the vehicle evasion control instruction is determined.

[0160] S209, it is determined whether the execution deviation exceeds a preset execution deviation threshold.

[0161] If yes, S210 is performed;

[0162] If no, S201 is performed.

[0163] S210, triggering re-planning of the vehicle evasion control instruction, and using a sample including the state feedback data and the execution deviation as incremental data, online updating model parameters of the graph neural network through an incremental learning algorithm to optimize accuracy of subsequent vehicle collision risk prediction.

[0164] In summary, the present application has at least the following advantages:

[0165] I. By constructing a dynamic interaction network and introducing a hierarchical attention refinement mechanism, the calculation overload and focal point dispersion problem of traditional attention models in high-density traffic flow is effectively improved. The mechanism extracts deep feature representations of each node through the graph neural network, and then quantitatively evaluates the influence score of each node, and further realizes adaptive attention sparsification screening based on a dynamic threshold and the influence score of each node, so as to accurately distinguish key risk vehicles from secondary interference targets, while ensuring real-time response, significantly improving the real-time, accuracy, and robustness of risk prediction, not only ensuring efficient processing within a limited calculation period, but also providing a highly reliable decision basis for the vehicle evasion control system by reducing the false positive rate and the false negative rate, thereby effectively ensuring the safety and reliability of the vehicle collision avoidance system in actual complex traffic environments.

[0166] II. Through the multi-layer iterative message passing mechanism of the graph neural network, the representation ability of the vehicle interaction features is effectively improved. This structure aggregates the node's own features and neighbor node features, enabling each node to integrate local topological relationship information. After linear transformation and activation function processing, a high-dimensional feature representation containing multiple levels of interaction context is gradually constructed. This hierarchical feature learning mechanism enables the model to capture complex spatio-temporal dependencies in traffic scenarios, laying a solid feature foundation for subsequent accurate assessment of node influence scores, thereby significantly enhancing the context awareness capability of risk prediction.

[0167] III. By establishing a closed-loop feedback mechanism from decision-making to execution, the real-time adaptability and long-term reliability of the system are effectively improved. By comparing the expected effect of the vehicle avoidance control instruction with the actual execution deviation of the vehicle state, the instruction can be re-planned immediately when control failure risk is detected, ensuring timely and effective execution of the avoidance action. At the same time, the execution deviation data is used as an incremental learning sample to optimize the graph neural network online, ensuring continuous learning and self-correction from actual interactions. This dual optimization mechanism not only guarantees the accuracy of single control response, but also improves the prediction accuracy and generalization ability in complex traffic scenarios by continuously accumulating practical experience.

[0168] IV. By enhancing the computing mechanism, the potential risk brought by the acceleration difference is considered in addition to the vehicle position and speed features. When the acceleration change direction of two vehicles is opposite or the difference is large, the interaction intensity is significantly improved, highlighting the impact of emergency dynamic changes on collision risk and enabling the edge weight of the dynamic interaction network to more sensitively reflect the risk propagation characteristics of sudden events in traffic flow. This allows the subsequent graph neural network to pay more attention to high-risk interaction relationships when aggregating features, further improving the response speed and accuracy of vehicle collision risk prediction in complex traffic environments.

[0169] V. Compared with existing technologies, the prediction delay or false alarm problem caused by computational resource overload and focus dispersion in traditional attention mechanisms in high-density traffic flow is effectively solved. The system can accurately and real-time identify the core collision risk source in extremely complex scenarios, dynamically balance the breadth and depth of attention according to the specific scene, and achieve adaptive improvement of the robustness and sensitivity of the vehicle collision risk prediction model under different traffic loads. This effectively breaks through the limitations of traditional open-loop or fixed parameter systems, enabling the entire system to continuously learn and evolve from actual interactions, thereby achieving stable optimization of long-term performance and high adaptability to changes in vehicle state and environment. In addition, the whole chain technical problems from perception, cognition to decision execution are systematically improved, ultimately achieving the core technical effect of significantly reducing the incidence of collisions and secondary accidents in highly dynamic and uncertain real traffic environments.

[0170] The following is an embodiment of the device of the present application, which can be used to execute the method embodiments of the present application. For details not disclosed in the device embodiments of the present application, please refer to the method embodiments of the present application.

[0171] Figure 3 A structural schematic diagram of a vehicle collision risk prediction device is provided for the exemplary embodiments of the present application. As shown in Figure 3 The vehicle collision risk prediction device 30 includes an acquisition module 31, a dynamic interaction network construction module 32, a feature extraction module 33, a determination module 34, a comparison module 35, and a risk prediction module 36, wherein:

[0172] The acquisition module 31 is configured to acquire running state data of the vehicle and road environment data.

[0173] The dynamic interaction network construction module 32 is configured to construct a dynamic interaction network based on the running state data and the road environment data, wherein the dynamic interaction network takes the vehicle and other traffic participants as nodes and the interaction relationship between the nodes as edges.

[0174] The feature extraction module 33 is configured to extract feature representations of the nodes in the dynamic interaction network based on a graph neural network.

[0175] The determination module 34 is configured to determine influence scores of the nodes based on the feature representations of the nodes, wherein the influence scores are used to quantify the influence degree of the nodes on the vehicle collision.

[0176] The comparison module 35 is configured to compare the influence scores of the nodes with dynamic thresholds respectively, and generate a sparse attention weight distribution according to the comparison results, wherein the dynamic thresholds are determined based on the running state data and the road environment data.

[0177] The risk prediction module 36 is configured to perform weighted fusion on the feature representations of the nodes based on the sparse attention weight distribution, to obtain a prediction result of the vehicle collision risk.

[0178] In one possible implementation, the feature extraction module 33 can be specifically configured to: perform iterative message passing on the dynamic interaction network through multiple network layers of the graph neural network; for each node in the nodes, update the feature representation of the node at the L+1 layer by: aggregating the feature representation of the node at the L layer and the feature representations of the neighbor nodes of the node at the L layer; performing linear transformation on the aggregated result, and processing the linear transformation result by using an activation function to generate the feature representation of the node at the L+1 layer; wherein L is a layer index.

[0179] In a possible implementation, the determining module 34 can be specifically configured to: multiply the norm of the feature representation corresponding to the node and the topological centrality metric value of the node in the dynamic interaction network, and take the multiplication result as the influence score of the node, where the norm of the feature representation corresponding to the node is used to quantify the risk urgency of the node causing direct collisions, and the topological centrality metric value is used to quantify the degree of the node as a key pivot for spreading collisions in the dynamic interaction network.

[0180] In a possible implementation, the running state data includes a vehicle speed of the ego vehicle, the road environment data includes a traffic flow density, and the dynamic threshold is determined by: inputting the traffic flow density and the vehicle speed of the ego vehicle into a preset linear relationship model to obtain a preliminary threshold; determining a consistency score and a false alarm rate for evaluating the accuracy of the prediction result of the vehicle collision risk based on historical actual traffic condition data and the prediction result of the vehicle collision risk within a preset setting time period; determining a reward signal based on the consistency score, the false alarm rate, and a preset penalty coefficient; and iteratively updating model parameters of the linear relationship model by using a policy gradient algorithm to maximize the reward signal, to perform online optimization on the preliminary threshold, and to generate the dynamic threshold.

[0181] In a possible implementation, the risk prediction module 36 can be specifically configured to: generate a vehicle avoidance control instruction based on the prediction result of the vehicle collision risk, and send the vehicle avoidance control instruction to an execution mechanism of the ego vehicle; collect state feedback data of the ego vehicle in real time, and determine an execution deviation between the state feedback data and expected state feedback data corresponding to the vehicle avoidance control instruction; if the execution deviation exceeds a preset execution deviation threshold, trigger re-planning of the vehicle avoidance control instruction; and meanwhile, take a sample including the state feedback data and the execution deviation as incremental data, and perform online updating on model parameters of the graph neural network by using an incremental learning algorithm, to optimize the accuracy of subsequent vehicle collision risk prediction.

[0182] In a possible implementation, the prediction result of the vehicle collision risk includes a risk probability value of the vehicle collision, a predicted risk time interval, a predicted risk spatial coordinate range, and a risk vehicle identification list, and the risk prediction module 36 can be further configured to: determine a target threat vehicle based on the risk vehicle identification list, determine a minimum safety distance required for the ego vehicle to achieve full braking within the predicted risk time interval based on a speed of the ego vehicle, the predicted risk time interval, and a maximum deceleration of the ego vehicle, and take the minimum safety distance as a safety spatial boundary; construct an optimization objective function with the safety spatial boundary as a hard constraint, the optimization objective function being configured to minimize the amplitude of the control instruction and optimize a ride comfort index within a future prediction time domain; obtain an optimal control sequence including braking, steering, and power control amounts by rolling solving the optimization objective function through a model predictive control algorithm; and output the control amounts corresponding to the current time in the optimal control sequence as the vehicle evasion control instruction.

[0183] In a possible implementation, in the step of constructing the dynamic interaction network, the weight of each edge in the dynamic interaction network is determined by: determining an interaction intensity based on the position vector and the velocity vector of any two nodes in the nodes; determining a relative acceleration length based on the acceleration vector of the two nodes; determining a scaling coefficient based on the relative acceleration length; multiplying the interaction intensity and the scaling coefficient, and taking the multiplication result as the weight of the edge between the two nodes; and wherein the scaling coefficient is positively correlated with the relative acceleration length, and the interaction intensity is amplified when the relative acceleration length increases.

[0184] The vehicle collision risk prediction apparatus provided by the embodiments of the present application can implement the technical solutions shown in the vehicle collision risk prediction method embodiments, and the implementation principles and beneficial effects are similar, and thus will not be described again.

[0185] It should be noted that, for each method embodiment described above, in order to simply describe, the method is described as a series of action combinations, but those skilled in the art should know that the present application is not limited to the action sequence described, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present application.

[0186] It should be noted that, although the steps in the flowchart are shown in a sequential order following the arrows, the steps are not necessarily executed in the order shown by the arrows. Unless otherwise specified herein, the steps are not necessarily executed in a strict order, and the steps can be executed in other orders. Moreover, at least some of the steps in the flowchart can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution of the sub-steps or stages can not necessarily be sequential, but can be performed in rotation or alternation with at least some of the other steps or sub-steps or stages of other steps.

[0187] It should be noted that the above-mentioned device embodiments are only illustrative, and the device of the present application can also be implemented in other ways. It should be understood that the division of the above device into various modules is only a logical functional division, and all or part of the modules can be integrated into one physical entity, or can be physically separated. The modules can all be implemented in the form of software called by a processing element, or all be implemented in the form of hardware, or part of the modules can be implemented in the form of software called by a processing element, and part of the modules can be implemented in the form of hardware. For example, the risk prediction module can be a separately established processing element, or can be integrated into a certain chip of the above device, or can be stored in the form of program code in the memory of the above device, and the function of the above risk prediction module can be called and executed by a certain processing element of the above device. The implementation of other modules is similar. In addition, all or part of the modules can be integrated together, or can be independently implemented. The processing element herein can be an integrated circuit having signal processing capability. In the implementation process, the steps of the above method or the above various modules can be completed by the integrated logic circuit of hardware or the instructions of software in the processing element.

[0188] For example, the above modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), or one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs), etc. For another example, when a certain module above is implemented by a form of processing element scheduling code, the processing element can be a general purpose processor, such as a Central Processing Unit (CPU) or other processor capable of invoking program code. For another example, the modules can be integrated together to be implemented in a form of System-On-a-Chip (SOC).

[0189] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present application are generated. The computer can be a general purpose computer, a special purpose computer, a computer network or other programmable device. The computer instructions can be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, Digital Subscriber Line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) mode. The computer readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available media sets. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a Digital Video Disc (DVD)) or a semiconductor medium (such as a solid state disk (SSD)) etc.

[0190] Figure 4 The structure of the controller provided in the exemplary embodiments of the present application is shown in the schematic diagram. As shown in the figure, Figure 4 The controller 40 of the present embodiment includes:

[0191] at least one processor 41; and a memory 42 connected with the at least one processor in communication;

[0192] The memory 42 stores instructions executable by the at least one processor 41, and the instructions are executed by the at least one processor 41 to enable the controller to perform the method of any of the above embodiments.

[0193] Optionally, the memory 42 can be independent or integrated with the processor 41.

[0194] The memory 42 can include a high-speed Random Access Memory (RAM) and can also include a non-volatile memory such as at least one disk memory.

[0195] The processor 41 can be a Central Processing Unit (CPU), or an Application Specific Integrated Circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application. Specifically, when implementing the vehicle collision risk prediction method described in the foregoing method embodiments, the controller can be, for example, a server or other electronic device with processing functions.

[0196] Optionally, the controller can further include a communication interface 43. In a specific implementation, if the communication interface 43, the memory 42 and the processor 41 are implemented independently, the communication interface 43, the memory 42 and the processor 41 can be connected with each other through a bus and complete communication therebetween. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc., but does not mean that there is only one bus or only one type of bus.

[0197] Optionally, in a specific implementation, if the communication interface 43, the memory 42 and the processor 41 are integrated on a chip, the communication interface 43, the memory 42 and the processor 41 can complete communication through an internal interface.

[0198] The implementation principle and technical effects of the controller provided in the embodiments can be referred to the foregoing embodiments, which will not be described here.

[0199] The embodiment of the present application further provides a computer program product, comprising a computer program which, when executed, implements the method steps in the method embodiments described above, and the specific implementation manners and technical effects are similar, and thus are not described herein.

[0200] The embodiment of the present application further provides a computer readable storage medium, which stores computer execution instructions, and when the computer execution instructions are executed, the computer execution instructions are used to implement the method steps in the method embodiments described above, and the specific implementation manners and technical effects are similar, and thus are not described herein.

[0201] The computer readable storage medium described above can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general or special purpose computer.

[0202] An exemplary readable storage medium is coupled to the processor, so that the processor can read information from the readable storage medium and write information to the readable storage medium. Of course, the readable storage medium can also be an integral part of the processor. The processor and the readable storage medium can be located in a special integrated circuit. Of course, the processor and the readable storage medium can also exist as discrete components in the vehicle collision risk prediction device.

[0203] The division of units is only a logical functional division, and in actual implementation, there can be another division manner, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0204] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, i.e., may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0205] In addition, each functional unit in various embodiments of the application can be integrated into one processing unit, or each unit can exist physically, or two or more units can be integrated into one unit.

[0206] If the function is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the various embodiment methods of the application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0207] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware. The aforementioned program can be stored in a computer readable storage medium. The program executes to perform the steps of the above-mentioned method embodiments; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk, and various media that can store program codes.

[0208] Finally, it should be noted that those skilled in the art, after considering the specification and practicing the application disclosed herein, will easily think of other embodiments of the application. The application is intended to cover any variations, uses, or adaptations of the application that follow the general principles of the application and include common knowledge or conventional techniques in the art that are not disclosed by the application, and is not limited to the precise structure described above and shown in the drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the application is only limited by the appended claims.

Claims

1. A method for predicting vehicle collision risk, characterized in that, include: Obtain the vehicle's operating status data and road environment data; Based on the operational status data and the road environment data, a dynamic interaction network is constructed; wherein, the dynamic interaction network uses the vehicle itself and other traffic participants as nodes, and the interaction relationships between the nodes as edges; Feature representations of each node in the dynamic interaction network are extracted based on graph neural networks; Based on the feature representation of each node, an influence score for each node is determined, and the influence score is used to quantify the degree of influence of the node on the occurrence of a vehicle collision. The influence score of each node is compared with a dynamic threshold, and a sparse attention weight distribution is generated based on the comparison results. The dynamic threshold is determined based on the running status data and the road environment data. Based on the sparse attention weight distribution, the feature representations of each node are weighted and fused to obtain the prediction result of vehicle collision risk.

2. The vehicle collision risk prediction method according to claim 1, characterized in that, The step of extracting feature representations of each node in the dynamic interaction network based on a graph neural network includes: The dynamic interaction network is iteratively message-passed through multiple network layers of the graph neural network. For each of the nodes, the feature representation of the node at layer L+1 is updated in the following way: Aggregate the node's own feature representation at layer L with the feature representations of each of the node's neighboring nodes at layer L; The aggregated result is subjected to a linear transformation, and the result is processed by an activation function to generate the feature representation of the node in the (L+1)th layer; where L is the layer index.

3. The vehicle collision risk prediction method according to claim 2, characterized in that, The determination of the influence score of each node based on the feature representation of each node includes: The norm of the feature representation corresponding to the node is multiplied by the topological centrality metric of the node in the dynamic interactive network, and the result of the multiplication is used as the influence score of the node. The norm of the feature representation corresponding to the node is used to quantify the urgency of the risk of the node causing a direct collision, and the topological centrality metric is used to quantify the criticality of the node in the dynamic interactive network in the propagation of collisions.

4. The vehicle collision risk prediction method according to claim 1, characterized in that, The operating status data includes the vehicle speed, the road environment data includes traffic flow density, and the dynamic threshold is determined in the following way: The traffic flow density and the vehicle speed are input into a preset linear relationship model to obtain a preliminary threshold. Based on the predicted results of vehicle collision risk over a historical time period and actual traffic data, a consistency score and false alarm rate are determined for the accuracy of the predicted results used to assess vehicle collision risk. A reward signal is determined based on the consistency score, the false alarm rate, and the preset penalty coefficient. With the goal of maximizing the reward signal, the model parameters of the linear relationship model are iteratively updated using a policy gradient algorithm to optimize the initial threshold online and generate the dynamic threshold.

5. The vehicle collision risk prediction method according to any one of claims 1 to 4, characterized in that, Also includes: Based on the predicted vehicle collision risk, a vehicle avoidance control command is generated and sent to the actuator of the vehicle. The status feedback data of the vehicle is collected in real time, and the execution deviation between the status feedback data and the expected status feedback data corresponding to the vehicle evasion control command is determined. If the execution deviation exceeds a preset execution deviation threshold, the vehicle evasion control command will be re-planned. Meanwhile, the samples containing the state feedback data and the execution deviation are used as incremental data, and the model parameters of the graph neural network are updated online using an incremental learning algorithm to optimize the accuracy of subsequent vehicle collision risk prediction.

6. The vehicle collision risk prediction method according to claim 5, characterized in that, The predicted vehicle collision risk includes a vehicle collision probability value, a predicted risk time interval, a predicted risk spatial coordinate range, and a list of risk vehicle identifiers. Based on the predicted vehicle collision risk, a vehicle avoidance control command is generated, including: Based on the risk vehicle identification list, target threat vehicles are identified, and based on the vehicle speed, the predicted risk time interval, and the vehicle's maximum deceleration, the minimum safe distance required for the vehicle to achieve full braking within the predicted risk time interval is determined, and the minimum safe distance is used as the safe space boundary. Using the safety space boundary as a hard constraint, an optimization objective function is constructed. The optimization objective function is configured to minimize the amplitude of the control command and optimize the ride comfort index within a future prediction time domain. By using a model predictive control algorithm, the optimization objective function is solved in a rolling manner to obtain the optimal control sequence that includes braking, steering and power control quantities; The control quantity corresponding to the current moment in the optimal control sequence is output as the vehicle evasion control command.

7. The vehicle collision risk prediction method according to any one of claims 1 to 4, characterized in that, In the step of constructing the dynamic interactive network, the weights of each edge in the dynamic interactive network are determined in the following way: For any two nodes among the nodes, the interaction strength is determined based on the position vector and velocity vector of the two nodes; Based on the acceleration vectors of the two nodes, the magnitude of the relative acceleration is determined; The scaling factor is determined based on the magnitude of the relative acceleration; Multiply the interaction strength by the scaling factor, and use the result of the multiplication as the edge weight of the two nodes; The scaling factor is positively correlated with the magnitude of the relative acceleration. When the magnitude of the relative acceleration increases, it amplifies the interaction intensity.

8. A vehicle collision risk prediction device, characterized in that, include: The acquisition module is used to acquire the vehicle's operating status data and road environment data; A dynamic interactive network construction module is used to construct a dynamic interactive network based on the operating status data and the road environment data; wherein, the dynamic interactive network uses the vehicle and other traffic participants as nodes, and the interaction relationships between the nodes as edges; The feature extraction module is used to extract feature representations of each node in the dynamic interaction network based on a graph neural network; The determination module is used to determine the influence score of each node based on the feature representation of each node, wherein the influence score is used to quantify the degree of influence of the node on the occurrence of a vehicle collision; The comparison module is used to compare the influence score of each node with a dynamic threshold, and generate a sparse attention weight distribution based on the comparison results, wherein the dynamic threshold is determined based on the running status data and the road environment data; The risk prediction module is used to perform weighted fusion of the feature representations of each node based on the sparse attention weight distribution to obtain the prediction result of vehicle collision risk.

9. A controller, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory is used to store computer-executed instructions; The processor is configured to execute the computer execution instructions to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed, are used to implement the method as described in any one of claims 1 to 7.

Images